topocast 0.0.4

• topocast() now brings every coarse grid a call fits -- each response's
  coefficients and diagnostics, plus the shared valid-cell count -- onto the target
  in one resample (grid target) or extract (point target), resolves the target
  predictors once, and assembles the coarse grids as a single multi-layer object
  rather than one terra object per coefficient. Each of those was previously done
  once per response, so a call repeated its whole coarse-to-target trip for every
  response it fit. The per-call fixed cost falls from about 0.58 s to about 0.11 s,
  a 25,000-point call with 19 responses from 1.28 s to 0.46 s, and the S4
  construction and dispatch that took over a fifth of a call no longer appears in
  its profile. With anomaly, the period stack likewise travels in one trip rather
  than one per period. Results are unchanged (#35).

• The moving-window engine's summed-area tables are now built and queried in
  double-double (extended) precision rather than plain double. A summed-area
  table entry near the far side of a large grid holds a sum over most of the
  grid's cells, so its magnitude grows with the grid size; on a plain-double
  table, a window's sum (extracted as the difference of four such entries) then
  loses precision proportional to that magnitude rather than to the window's own
  size, which by roughly a million cells could already turn a genuinely zero
  within-window variance into a small positive number that wrongly cleared
  min_variance, defeating the no-spread guard (#21).

• arma::uword is now explicitly 64-bit (ARMA_64BIT_WORD), and the moving
  window's row/column bounds are computed in a 64-bit type, closing two overflow
  classes left open after #13/#14 for a grid whose total cell count or single
  dimension is large enough to overflow a 32-bit computation (#31, #32).

• The C++ engine calls Rcpp::checkUserInterrupt() between chunks of rows (and
  during single-threaded setup), so a long-running fit on a large grid can now be
  interrupted rather than only checked for completion (#33).

• A response or predictor named r.squared, residual.sd, n.valid, or
  (Intercept) is now rejected at the formula-parsing stage; those names are
  reserved for topocast()'s own diagnostic and coefficient columns, and a
  response sharing one previously had its downscaled values silently overwritten
  by the diagnostic column of the same name when diagnostics = TRUE (#22).

• n.valid is now bounded to [0, (2 * radius + 1)^2] after resampling, the
  true range a window's valid-cell count can hold; like r.squared, it could
  previously be carried slightly outside its true range by some resampling
  methods (#24).

• radius = 0 is now rejected: its one-cell window can never hold the k + 1
  valid cells a fit with k predictors needs, so it previously returned an
  all-NA result with no error or warning (#25).

• An invalid aggregate or grid-target method now errors with a message
  naming the actual topocast() argument, instead of terra's raw
  'arg' should be one of ... message, which named neither (#26).

• With type = "ratio", a coarse cell where the baseline and the anomaly period
  are both exactly zero (e.g. a dry-season precipitation cell with no rain in
  either) is now treated as ratio 1, "no change", rather than discarding the
  fine baseline the same way a true x / 0 does (#27).

• A CRS mismatch error now falls back to a PROJ4 or WKT snippet when neither
  side carries an identifiable EPSG code (common for a custom or regional
  projection), instead of printing "unknown CRS" on both sides with no way to
  tell how they actually differ (#28).

• An entirely non-finite response/predictor stack now errors with a message
  naming the actual inputs (data/onto for topocast(), y/x for
  window_regression()) instead of the C++ engine's generic
  "no finite cells to regress" (#30, #34).

• output = "raster" no longer silently renames the (Intercept) coefficient
  column via the raster package's own make.names()-based names<-, which is
  now documented as an inherent limitation of that output class rather than left
  unmentioned (#23).

• inst/CITATION now reports the installed package version dynamically instead
  of a hardcoded string that had drifted out of date (#29).

• Several responses that share the predictors are downscaled in one call by naming
  them as cbind(prec, tmin) ~ elev. The moving-window design depends only on the
  predictors, so it is assembled and factored once and solved against every response;
  each extra response adds only a back-substitution. The result has one layer (or
  column) per response, and coefficients/diagnostics grids are prefixed by the
  response name. window_regression() likewise accepts a list of response matrices.

• diagnostics = TRUE returns an r.squared grid: the per-window coefficient of
  determination of the local fit, computed from the same summed-area sufficient
  statistics, mapping where the terrain relationship is strong. window_regression()
  now returns r_squared alongside intercept and slope.

• diagnostics = TRUE also returns residual.sd (the residual standard deviation of
  the local fit, in the response's own units) and n.valid (the count of valid
  coarse cells the window held), from the same sufficient statistics as r.squared.
  n.valid is shared across responses rather than prefixed, since the valid-cell mask
  is complete-case across them; residual.sd is prefixed like r.squared is.
  window_regression() likewise returns residual_sd and n_valid (#8).

• clamp = TRUE bounds the downscaled field to the observed range of the coarse
  response, a guard against the local linear fit extrapolating without limit where a
  fine predictor lies outside the range it was fit on.

• The per-cell window solve runs in parallel with OpenMP where the toolchain
  provides it; results are unchanged.

• anomaly and baseline accept a Raster* (raster) or stars object, like data
  and onto already did, instead of only a SpatRaster (#5).

• radius and min_cells are validated as non-negative whole numbers before
  reaching the C++ engine. A negative radius previously indexed the summed-area
  table out of bounds and crashed the R session rather than erroring (#4); the same
  validation now also rejects a value too large for a 32-bit integer, which
  previously overflowed to NA on coercion and crashed the session the same way (#9).

• min_variance is validated as a non-negative number. NA or a negative value
  previously compared as false against every window variance, silently disabling
  the documented guard against a predictor with no spread instead of being
  rejected (#10).

• window_regression() rejects a partially-named list of responses instead of
  silently making the unnamed ones' results unreachable by name (#6).

• window_regression() rejects a list of responses with a repeated name, which
  had the same unreachable-by-name problem as #6 but was not caught by it (#15).

• A min_cells close to .Machine$integer.max no longer overflows the C++
  engine's valid-cell threshold; previously the overflow silently disabled the
  guard instead of enforcing it, so windows with far fewer than the requested
  minimum could still be fit (#13).

• A radius close to .Machine$integer.max no longer crashes the R session;
  it is now clamped to the grid's span, which covers every cell in every
  window anyway and so changes no result (#14).

• diagnostics = TRUE's n.valid (and window_regression()'s n_valid) now
  reports a window's valid-cell count whenever the window held enough cells to
  attempt a fit, rather than only when the fit also succeeded; a degenerate
  (no-spread predictor) or singular window previously reported n.valid = NA
  indistinguishably from a window that never had enough cells (#16).

• A window whose predictors are exactly collinear (rank-deficient design) now
  reliably returns NA instead of a spurious least-squares coefficient;
  Armadillo's default solver silently falls back to an approximate solution on
  a detected-singular system rather than failing, which defeated the
  documented singular-design guard (#17).