Mask 64-bit integer sentinels at native width in eager reads (#3098)

[brendancol]

Closes #3098

• _apply_eager_nodata_mask compared the integer nodata sentinel after promoting the buffer to float64. For sentinels above 2**53 (INT64_MAX, UINT64_MAX), float64 rounding made up to 512 (int64) or 1024 (uint64) nearby valid values compare equal to the sentinel, so eager masked reads turned them into NaN. The fix computes the mask at the source dtype's width before the promotion, the same order the dask per-chunk mask, the GPU GDS chunk path, and the VRT path already use.
• The unconditional float64 promotion from geotiff: eager vs lazy nodata masking diverge on dtype and masked_nodata for int sources with no sentinel pixels #2990 is unchanged, and the mask now agrees with the native-width nodata_pixels_present scan in the same function.
• One site fixes both eager backends: the GPU eager read calls the same helper on CuPy buffers via duck typing.

Backend coverage: numpy and cupy eager were affected and are fixed; dask+numpy and dask+cupy were already correct and now match the eager output (verified on hardware with CUDA).

Test plan:

• 5 regression tests in tests/read/test_nodata.py: int64 exact-hit masking, eager-vs-dask NaN-pattern parity, uint64 variant, near-sentinel file with no exact hit (nodata_pixels_present False, still promoted to float64), and the cupy eager path
• tests/read/ + tests/unit/: 2199 passed, 5 skipped
• tests/parity/test_finalization.py and tests/test_round_trip.py pass
• flake8 clean on the touched files

[brendancol]

PR Review: Mask 64-bit integer sentinels at native width in eager reads (#3098)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• xrspatial/geotiff/tests/read/test_nodata.py:3517 (test_eager_masked_int64_near_sentinel_without_hit_3098 docstring): "the native\n-width scan" wraps mid-word and reads as "native -width". Rejoin as "native-width".

What looks good

• The fix is the minimal reorder: compute the equality mask before astype(np.float64), leave the unconditional promotion from #2990 alone. For dtypes of 32 bits and below every value is exact in float64, so behavior is unchanged there; only the broken 64-bit case moves.
• It also resolves the internal inconsistency where the mask_nodata=False scan in the same function compared at native width but the masking branch did not, so nodata_pixels_present and the actual mask could disagree.
• One site covers both affected backends. The GPU eager read reaches this helper via duck typing from _finalize_eager_read, and all the ops involved (arr == scalar, boolean-mask assignment, astype) were already exercised on CuPy by the previous code, just in a different order. The GPU regression test confirms it on hardware.
• Tests pin the four interesting cases: exact-hit-only masking for INT64_MAX, eager-vs-dask NaN-pattern parity on the same file, the wider UINT64_MAX window, and the no-exact-hit file where nodata_pixels_present must be False while the buffer still promotes to float64.

Checklist

• Algorithm matches reference (dask _delayed_read_window, _apply_nodata_mask_gpu, and the VRT mask all use the same native-width order)
• All implemented backends produce consistent results (verified numpy, cupy, dask+numpy, dask+cupy with CUDA)
• NaN handling is correct (only exact sentinel hits become NaN)
• Edge cases are covered by tests (near-sentinel values, no-hit file, uint64 max)
• Dask chunk boundaries handled correctly (no dask code changed; parity test reads with chunks=2 across the affected pixels)
• No premature materialization or unnecessary copies (mask now computed once on the int buffer instead of the promoted one)
• Benchmark not needed (mask is a single vectorized compare either way)
• README feature matrix not applicable (bug fix, no API change)
• Docstrings present and accurate (helper comment explains the 2**53 rounding and cites #3098)

[brendancol]

Follow-up review after f939219: the docstring nit is fixed and the 5 regression tests still pass. No remaining findings. Disposition of the original review: 0 blockers, 0 suggestions, 1 nit (fixed in-PR).