Refresh nodata/nodatavals on resample identity fast path (#2662)

[brendancol]

Closes #2662

What changed

• The resample identity fast path (scale_factor=1.0) and the 3D dispatch path masked nodata sentinels to NaN but refreshed only _FillValue, leaving attrs['nodata'] and attrs['nodatavals'] pointing at the original finite sentinel. Output advertised nodata=-9999 while the data was already NaN.
• Factored the non-identity path's refresh logic into a _refresh_nodata_attrs helper and called it from all three paths, so masked-to-NaN output never advertises a stale integer sentinel.

Backend coverage

The fix is backend-independent attr handling; the mask itself routes through xarray.where. Regression tests cover numpy and dask+numpy. cupy and dask+cupy masking is exercised by the existing cross-backend nodata coverage.

Test plan

• New TestIdentityNodataMetadata class covers: identity path refreshes all three attrs, identity matches the non-identity path, absent attrs stay absent, and the 3D path is fixed too.
• Full resample suite passes (289 tests).

[brendancol]

PR Review: Refresh nodata/nodatavals on resample identity fast path (#2662)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• xrspatial/resample.py identity path (~L1392) and non-identity path (~L1503): the explicit _FillValue = float('nan') line is redundant whenever _FillValue is already in the input attrs, since _refresh_nodata_attrs sets it too. It is still needed for the explicit-nodata=-only case where no _FillValue attr exists, so leaving it is correct. A one-line comment noting that would help the next reader, but it is not required.

What looks good

• The fix is symmetric across the three code paths: identity, 3D, and non-identity now all run if has_nodata: _FillValue=nan followed by _refresh_nodata_attrs, so they agree on what the output advertises.
• The shared helper reads from the original src_attrs and writes to dst_attrs. That avoids the order-dependence bug you would hit if it read the partially-updated output attrs instead.
• Tests hit the regression directly: masked pixel is NaN and no attr advertises the stale sentinel, identity matches the non-identity path, absent attrs stay absent, and the 3D path is covered.
• The cupy/dask+cupy parametrize was dropped with a documented reason. The attr fix is backend-independent, so numpy + dask+numpy is enough.

Checklist

• Algorithm matches intended behavior
• Backends produce consistent results (attr fix is backend-independent)
• NaN handling is correct
• Edge cases covered (absent attrs, 3D path)
• Dask chunk boundaries: not applicable, no map_overlap change
• No premature materialization or unnecessary copies
• Benchmark: not needed, no perf-sensitive change
• README feature matrix: not applicable, no new function
• Docstrings present and accurate

[brendancol]

Follow-up review (#2662)

The one nit from the first pass is addressed: the identity path now carries a comment explaining that the explicit _FillValue = nan line covers the explicit-nodata=-only case (where the input had no nodata attrs), while _refresh_nodata_attrs handles inputs that declared nodata / nodatavals. Logic is unchanged.

Also merged origin/main into the branch. The merge was clean (no conflicts); git auto-merged resample.py and test_resample.py. The full resample suite passes (222 tests) after the merge.

No new blockers, suggestions, or nits.