resample: refresh transform to match actual coord orientation (#2571)

[brendancol]

Closes #2571.

Summary

• resample() was always writing a north-up rasterio transform
  (+x, -y, top-left origin), even when the input had ascending y
  or descending x coords. The refreshed attrs['transform'] then no
  longer matched the coords on the same DataArray, so downstream
  consumers that georeference via transform would see a flipped
  raster.
• Use the signed px / py from _new_coords and anchor the
  origin at *_edge_start (the leading edge on the side of
  vals[0]). This matches rasterio's (col=0, row=0) -> first array pixel convention for every coord orientation.

Backend coverage

The transform refresh runs in pure Python on the metadata after the
backend dispatch, so the fix applies to numpy, cupy, dask+numpy, and
dask+cupy identically. No backend-specific code changed.

Test plan

• pytest xrspatial/tests/test_resample.py::TestMetadataPropagation
  -- 19 passed, includes the 4 new orientation parametrizations.
• pytest xrspatial/tests/test_resample.py -- 182 passed.

[brendancol]

PR Review: resample transform refresh (#2571)

Blockers

None.

Suggestions

None.

Nits

• xrspatial/tests/test_resample.py:163 -- the _raster_with_orientation
  docstring says "upper-left edge of pixel [0, 0]", but for ascending
  y the first pixel sits at the geographic lower-left. The
  transform is still right; the wording just borrows rasterio's
  "upper-left" convention where "leading-edge corner of pixel [0, 0]"
  would be more honest.
• xrspatial/tests/test_resample.py:191 -- assert b == 0.0 and d == 0.0
  packs two facts into one assertion. Splitting them tells you which
  off-diagonal drifted when it fails.

What looks good

• The math holds up. px / py out of _new_coords are already
  signed, and *_edge_start is the leading edge on the vals[0]
  side, so transform * (col, row) lands where the output coords
  say it should regardless of orientation.
• The pre-existing test_transform_refreshed_on_downsample is
  untouched and still passes, which is a useful regression anchor.
• Metadata-only change, so the fix covers numpy, cupy, dask+numpy,
  and dask+cupy without any per-backend code.
• The parametrized test sweeps all four x asc/desc x y asc/desc
  combos and round-trips transform * (col+0.5, row+0.5) against
  the output pixel centers.

Checklist

• Algorithm matches the rasterio Affine convention
• All backends consistent (metadata-only)
• NaN handling unchanged
• Orientation edge cases covered
• Dask chunks N/A
• No premature materialization
• Benchmark not needed
• README feature matrix not applicable
• Docstrings unchanged

[brendancol]

Follow-up review (#2571)

Both nits from the previous review are fixed in 34dc110:

• _raster_with_orientation docstring now says "leading-edge corner"
  instead of "upper-left edge".
• The combined off-diagonal assert is split into two.

No new findings on the follow-up pass. The 19
TestMetadataPropagation tests still pass locally.