Keep integer dtype for empty reproject chunks (#3096)

[brendancol]

Closes #3096

• Empty-chunk fills in _reproject_chunk_numpy, _reproject_chunk_cupy, and the footprint-skip path in _reproject_block_adapter now use the source's integer dtype instead of a hardcoded float64. One no-overlap chunk used to promote the whole computed dask array to float64 while the lazy array (and the eager backend) advertised int16/uint8/etc. Float sources keep float64 fills.
• The resolved nodata is always representable for integer rasters (reproject: integer rasters silently corrupt nodata when default NaN is used #2185/reproject: out-of-range explicit nodata corrupts integer output and contradicts attrs #2572), so the integer fill is safe.
• Adds TestEmptyChunkDtype: dask+numpy int16 with forced empty chunks (advertised == computed == int16, sentinel in the empty corners), eager numpy with a fully no-overlap output grid, a float64 guard, and a dask+cupy variant.
• Also carries the accuracy-sweep state CSV update for the reproject module.

Backend coverage: numpy, cupy, dask+numpy, dask+cupy (the cupy worker and block adapter paths got the same fix).

Test plan:

• pytest xrspatial/tests/test_reproject.py -k EmptyChunkDtype (4 passed on a GPU host)
• pytest xrspatial/tests/test_reproject.py (436 passed)
• pytest xrspatial/tests/test_reproject_coverage_2026_05_27.py xrspatial/tests/test_reproject_cupy_gate_2564.py (12 passed)

[brendancol]

PR Review: Keep integer dtype for empty reproject chunks (#3096)

Blockers (must fix before merge)

None found.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• In _reproject_block_adapter, the 2-D footprint-skip path on dask+cupy still returns a numpy fill while the data chunks are cupy arrays (xrspatial/reproject/__init__.py:2003, pre-existing; the 3-D branch does return cp.full). The dask meta is numpy either way, so the stitched result behaves, but the array-type mix is worth a follow-up note if dask+cupy empty chunks ever grow a device-side consumer. Out of scope here since this PR only changes the fill dtype.
• A non-integral float nodata on an integer raster (e.g. explicit nodata=0.5 with int16) would truncate silently in the new np.full(..., dtype=int16) fills. The data path's round-and-clip cast already collapses such a sentinel the same way, so this PR doesn't make it worse, and _detect_nodata range-checks but doesn't integrality-check. Pre-existing input-validation gap, not introduced here.

What looks good

• All eleven empty-fill sites got the same treatment: four in _reproject_chunk_numpy, five in _reproject_chunk_cupy, and the footprint-skip path in _reproject_block_adapter, so the fix can't be dodged by which early-return a chunk happens to hit.
• The dtype rule matches the template logic in _reproject_dask and the eager cast-back exactly (np.issubdtype(..., np.integer)), so bool and float sources keep float64 fills and nothing else shifts.
• Leaning on the #2185/#2572 guarantee that integer rasters always resolve an in-range nodata is the right call; the integer fill is safe by construction and the comment says why.
• Tests pin the contract from four angles: dask advertised-vs-computed dtype with forced empty chunks (plus a sentinel check in the empty corner and a has-real-data guard so the test can't pass vacuously), the eager fully-no-overlap case, a float64 regression guard, and a dask+cupy run.
• Merge paths are unaffected: the canonicalize step already promotes via np.where(arr == r_nd, np.nan, arr), which handles integer fills the same as the old float64 ones.

Checklist

• Algorithm unchanged (fill dtype only; values identical)
• All implemented backends produce consistent results (numpy, dask+numpy, cupy worker, dask+cupy covered)
• NaN handling is correct (float sources keep NaN fills, guarded by test)
• Edge cases are covered by tests (empty corners, fully-no-overlap output, float guard)
• Dask chunk boundaries handled correctly (per-chunk fills now match the template dtype)
• No premature materialization or unnecessary copies
• Benchmark not needed (no hot-path change; fills were already allocated)
• README feature matrix unchanged (no API or backend-support change)
• Docstrings/comments explain the why, with issue references

[brendancol]

Follow-up review (after 6a24085)

The first nit from the previous review is addressed: the 2-D footprint-skip path in _reproject_block_adapter now returns cp.full on dask+cupy, matching the 3-D branch and the data chunks. Full reproject suite re-run on a GPU host: 436 + 12 passed.

Remaining item, dismissed with reason: the non-integral-float-nodata-on-integer-raster truncation is a pre-existing validation gap in _detect_nodata (unmodified by this PR), and the data path's round-and-clip cast already collapses such a sentinel identically. No change needed here.

No new findings.