Preserve input float dtype in focal.mean() across all backends (#3214)

[brendancol]

Closes #3214

• mean() now uses _promote_float on every backend: float64 stays float64, float32 stays float32, integers promote to float32. Previously numpy/dask+numpy always returned float64 and cupy/dask+cupy always returned float32, so the output dtype depended on the backend.
• Drops the hardcoded cupy.float32 casts in _mean_cupy / _mean_dask_cupy. float64 rasters keep their precision on GPU; with values near 1e7 the old float32 path's error (0.58) was larger than the spread of the true focal means (0.42).
• excludes values are cast to the working dtype before matching, so exclusion behaves the same on CPU and GPU for float32 rasters.

This is the same contract apply() and focal_stats() got in #2769 and convolve_2d() got in #1096; mean() was the holdout.

Backend coverage: numpy, cupy, dask+numpy, dask+cupy. All four ran locally on a CUDA host.

The diff also updates one row of .claude/sweep-accuracy-state.csv (accuracy-sweep bookkeeping for the focal module).

Test plan:

• New dtype tests: float64 preserved, float32 preserved, int promotes to float32, parametrized across backends
• New GPU large-offset parity test (numpy vs cupy at offset 1e7, rtol 1e-12)
• New excludes-dtype matching test
• Full xrspatial/tests/test_focal.py: 245 passed on a GPU host
• test_accessor.py + test_emerging_hotspots.py: 113 passed

[brendancol]

PR Review: Preserve input float dtype in focal.mean() across all backends (#3214)

Blockers (must fix before merge)

None found.

Suggestions (should fix, not blocking)

• test_mean_excludes_match_in_working_dtype_3214 (xrspatial/tests/test_focal.py) only runs on numpy. The cupy path has its own excludes cast (cupy.asarray(excludes, dtype=fdtype) in _mean_cupy, xrspatial/focal.py:262-264), so the float32 exclusion-match behavior should be asserted there too. Parametrizing the test over numpy and cupy covers it.
• This PR changes integer-input behavior on the CPU backends: int rasters used to come back float64 and now come back float32, so int values above 2^24 are no longer exact in the output. That matches the apply()/focal_stats()/convolve_2d() contract, but the mean() docstring (xrspatial/focal.py, Returns section) says nothing about output dtype. One sentence stating the dtype contract would prevent surprise.

Nits (optional improvements)

• The two F401 'cupy' imported but unused warnings at test_focal.py:1688 and 1704 predate this PR (from the #2730 tests) and are correctly left alone here.

What looks good

• All four backends now share one dtype rule, applied at the dispatch boundary (mean()) and inside the chunk function (_mean_cupy), so the dask+cupy path is covered both lazily and per chunk.
• Casting excludes to the working dtype closes a cross-backend mismatch that the float32 change would otherwise have introduced (CPU comparing float32 cells against float64 excludes).
• The large-offset GPU test pins parity at rtol 1e-12 and asserts the focal-mean spread survives, which is the actual failure mode from the issue, not just a dtype check.
• A time_mean benchmark already exists in benchmarks/benchmarks/focal.py, so no benchmark gap.

Checklist

• Algorithm matches reference (no algorithm change; dtype handling only)
• All implemented backends produce consistent results (verified on a CUDA host)
• NaN handling is correct (NaN excludes survive the dtype cast; _equal_numpy uses isnan)
• Edge cases are covered by tests (float64/float32/int inputs, large offset, excludes matching)
• Dask chunk boundaries handled correctly (no depth changes; astype is lazy)
• No premature materialization or unnecessary copies
• Benchmark exists
• README feature matrix unchanged (no backend support change)
• Docstring does not state the output dtype contract (see suggestion above)

[brendancol]

Follow-up review (commit 83be5d1)

Both suggestions from the first pass are addressed:

• test_mean_excludes_match_in_working_dtype_3214 is now parametrized over numpy and cupy and handles the device-array extraction with .get(). Verified both variants pass on a CUDA host.
• The mean() Returns section now states the dtype contract (float preserved, int promotes to float32, matching apply and focal_stats).

No new findings. The pre-existing F401 nits remain out of scope.

Disposition of first-pass findings: 2 suggestions fixed, 1 nit dismissed (pre-existing, untouched by this PR), 0 blockers.