Surface standalone docstrings on .xrs accessor methods; fix broken hydro delegations

[brendancol]

Closes #2981

help(arr.xrs.slope) showed nothing useful because the .xrs accessor methods are thin delegating wrappers with no docstrings of their own. This copies each standalone function's docstring onto its accessor method at import time, so help(arr.xrs.slope) now matches help(slope).

While adding that, I found 17 hydrology accessor methods were broken: they imported from per-algorithm modules (from .fill import fill, .watershed, and so on) that had been consolidated into xrspatial/hydro.py, so calling them raised ModuleNotFoundError. Nothing in the accessor tests called them, so it had gone unnoticed. This PR repoints those delegations at .hydro too.

Changes

• Repoint the 17 broken hydrology delegations (fill, flow_direction, flow_accumulation, watershed, basin, basins, sink, stream_order, stream_link, snap_pour_point, flow_path, flow_length, twi, hand, plus the dinf/mfd variants) at xrspatial.hydro, in both the DataArray and Dataset accessors.
• Attach standalone docstrings to accessor methods. Methods that already had hand-written docs (plot, to_geotiff, open_geotiff) are left alone. The hydrology unified wrappers only carry a generic dispatcher stub, so their help text is taken from the documented default-algorithm *_d8 variants.

Doc-only — no __signature__ is set, since several methods (zonal_stats, crop, idw/kriging/spline, mahalanobis) don't pass the accessor's array as the function's first argument, which would make a naive signature wrong. The copied Parameters section already documents everything.

Backends

No backend-specific code. The docstring attachment is pure Python metadata, and the hydrology import fix routes through the same xrspatial.hydro functions that already support numpy / cupy / dask+numpy / dask+cupy.

Test plan

• pytest xrspatial/tests/test_accessor.py (64 passed)
• Docstring surfacing matches the source function (slope, aspect, hillshade, curvature, focal_mean->mean, ndvi, fill->fill_d8, watershed->watershed_d8)
• Drift guard: every public method on both accessor classes carries a non-stub docstring
• Regression: all 17 hydrology delegations resolve; the 6 single-input ones match their standalone function
• flake8 clean on changed files

Not touched: no new public API, so no user-guide notebook, README matrix row, or Sphinx reference page. The accessor has no reference page today; that's a separate docs task.

[brendancol]

PR Review: Surface standalone docstrings on .xrs accessor; fix broken hydro delegations

Reviewed the full diff and both changed files in the worktree. No blockers. The hydro import fix is verified end-to-end and the docstring mechanism is sound. A couple of robustness items in the new code are worth tightening.

Suggestions (should fix, not blocking)

• xrspatial/accessor.py, _delegated_doc: the lookup getattr(getattr(xrspatial, source_name, None), '__doc__', None) has a sharp edge. If a doc-source name is ever missing (a typo in _DOC_SOURCE_OVERRIDES, or a future rename), the inner getattr returns None, and None.__doc__ is the truthy string "The type of the None singleton." — which then gets attached as the method's docstring. It doesn't trigger today (every name resolves, drift guard is green), but it fails silently and wrong rather than skipping. Resolve the function first and read __doc__ only when it's not None: func = getattr(xrspatial, source_name, None); return func.__doc__ if func is not None else None.

• Dataset accessor multispectral methods (ndvi, evi, savi, etc.) now show the standalone function's docstring, which documents array inputs (nir_agg, red_agg). The Dataset accessor actually takes band-name strings (ds.xrs.ndvi(nir='b8', red='b4')), so the surfaced parameter docs don't match the call convention. The class docstring already explains the band-mapping, so this is informative-but-imperfect rather than wrong. Worth a note, or leave as-is given the goal was parity with the standalone help.

Nits (optional)

• xrspatial/accessor.py: the module-level for _cls in (...) loop leaves _cls bound in the module namespace. Harmless; del _cls after the loop or a comprehension would keep it clean.
• xrspatial/tests/test_accessor.py: import inspect and the accessor-class imports sit mid-file with # noqa: E402. Fine and flake8-clean, but they could move to the top with the other imports.

What looks good

• The hydro fix is the right root cause: the 17 methods imported from per-algorithm modules that were folded into xrspatial.hydro, and repointing the lazy imports there fixes all of them. Verified no remaining broken imports and that the six single-input methods match their standalone results.
• Docstrings are sourced from the documented *_d8 variants for the hydro wrappers, avoiding the generic dispatcher stub.
• Tests compare against real reference functions (not just "it runs"), and the drift guard will catch any future accessor method added without a usable docstring.
• Hand-written docs on plot/to_geotiff/open_geotiff are correctly left untouched.

Checklist

• Behavior matches intent (help text parity; hydro calls succeed)
• No backend-specific code; hydro fix routes through existing multi-backend functions
• No premature materialization or copies
• Edge cases covered by tests (delegation resolves, doc surfacing, drift guard)
• Docstrings present and accurate (one Dataset-multispectral caveat above)
• Latent None.doc fallback worth tightening (Suggestion 1)

[brendancol]

Follow-up review (after fixes in 01a5348)

Re-read the diff after the follow-up commit. The earlier findings are resolved; nothing new surfaced.

Disposition of the first-pass findings

• Suggestion 1 (None.doc fallback) — fixed. _delegated_doc now resolves the source function first and returns None when it's missing, so a bad override name no longer attaches "The type of the None singleton.". Verified with a bogus override.
• Nit 1 (_cls loop leak) — fixed. Replaced with two explicit _attach_delegated_docs(...) calls.
• Nit 2 (mid-file test imports) — fixed. inspect and the accessor-class imports moved to the top of the test module.
• Suggestion 2 (Dataset multispectral docstrings describe array inputs, not the band-name-string API) — kept as-is. Sourcing the standalone docstring is the stated goal, the Dataset accessor class docstring already explains the band mapping, and bespoke per-method Dataset docs would be a larger change for marginal gain. Tracking it as a known imperfection rather than a fix here.

Tests: 64 passed, flake8 clean on both changed files.