Preserve input attrs in mcda.constrain() (#3147)

[brendancol]

Closes #3147

• constrain() lost all input attrs (res, crs, nodatavals) whenever exclude was non-empty: xr.where(mask, fill, result) takes attrs from the scalar fill. With exclude=[] attrs survived, so behavior also depended on whether any mask was passed.
• The fix restores suitability.attrs (as a copy) on the output after the masking loop.
• New TestConstrainAttrs covers attrs survival with a mask, with a custom fill, with an empty exclude list, attrs-dict isolation from the input, and the dask+numpy and dask+cupy paths.

Backend coverage: numpy, dask+numpy, and dask+cupy verified end-to-end. On bare cupy, xr.where itself raises before the fix is reached (known cupy 13.6 / xarray 2025.12 incompatibility, tracked separately); the dask+cupy test asserts attrs on the lazy result for the same reason.

Test plan:

• pytest xrspatial/tests/test_mcda.py (181 passed, includes 6 new attrs tests)
• Manual 4-backend probe: attrs now kept on numpy / dask+numpy / dask+cupy

Found by the metadata propagation sweep. The sweep state CSV update for the mcda module rides along in this PR.

[brendancol]

PR Review: Preserve input attrs in mcda.constrain() (#3147)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• xrspatial/mcda/constrain.py:30-33: the Returns section could state that the output keeps the input's attrs. The fix establishes that contract; documenting it makes the next regression visible in review rather than in a sweep.

What looks good

• The root cause is identified precisely: xr.where takes attrs from its first value argument, the scalar fill, so any non-empty exclude stripped attrs. Restoring dict(suitability.attrs) after the loop fixes every mask count and fill value in one place, and the dict copy keeps the output from aliasing the input's attrs (covered by test_attrs_not_shared_with_input).
• Test coverage matches the bug shape: mask present, custom fill, empty exclude list, dask+numpy with value parity against numpy, and dask+cupy asserting attrs on the lazy result. The comment in test_attrs_kept_dask_cupy explaining why values are not computed there (pre-existing cupy/xarray xr.where incompatibility) is accurate; I reproduced that failure on main without this change.
• Hand-rolled fixtures match the existing test_mcda.py style, which does not use general_checks helpers anywhere.

Checklist

• Algorithm matches reference/paper (N/A, metadata-only fix)
• All implemented backends produce consistent results (numpy, dask+numpy verified by value; dask+cupy attrs verified lazily; bare cupy blocked upstream before this code runs)
• NaN handling is correct (unchanged; NaN fill still propagates)
• Edge cases are covered by tests (empty exclude, custom fill, attrs aliasing)
• Dask chunk boundaries handled correctly (no chunk-level code touched)
• No premature materialization or unnecessary copies
• Benchmark exists or is not needed (not needed, no compute change)
• README feature matrix updated (not applicable)
• Docstrings present and accurate (see nit: attrs contract not documented)

The sweep state CSV row riding along is expected for metadata-sweep PRs.

[brendancol]

Follow-up review after 9f-commit addressing the docstring nit:

• The Returns section of constrain now states the output keeps the input's dims, coords, and attrs. That closes the only finding from the first pass.
• No new code paths were touched; constrain tests still pass (13 passed locally).

Nothing outstanding.