geotiff: forward read kwargs through all GPU CPU-fallback paths

[brendancol]

Closes #2238.

Summary

• Forward allow_rotated, window, band, and max_pixels through every CPU-fallback _read_to_array call inside read_geotiff_gpu (stripped-layout, planar=2 per-band, sparse-tile, and post GPU-decode-failure).
• Skip the post-decode shape check and _gpu_apply_window_band slicer on CPU-decoded paths so the now-already-sliced fallback buffer doesn't get double-windowed or trip the multi-band shape guard.
• Flip test_cupy_rotated_read_drops_crs from xfail to pass; add a regression-test file pinning the kwarg-forwarding contract on all four fallback sites.

Backend coverage

Touches only xrspatial/geotiff/_backends/gpu.py. The cupy and dask+cupy paths use this code; the dask+cupy chunked path already forwarded kwargs correctly via _read_geotiff_gpu_chunked, so it's unaffected. The numpy and dask+numpy paths are unchanged.

Test plan

• pytest xrspatial/geotiff/tests/test_gpu_fallback_forwards_kwargs_2238.py (6 new tests, all pass)
• pytest xrspatial/geotiff/tests/test_allow_rotated_no_crs_2122.py (xfail-flipped test now passes; 8 tests total)
• pytest -k "gpu or cupy or planar or sparse or fallback or rotated" xrspatial/geotiff/tests/ (1189 passed, 20 skipped, no regressions)

[brendancol]

PR Review: geotiff: forward read kwargs through all GPU CPU-fallback paths

Blockers (must fix before merge)

• None.

Suggestions (should fix, not blocking)

• test_sparse_tile_fallback_respects_caller_max_pixels (test_gpu_fallback_forwards_kwargs_2238.py:351-376) does not actually exercise the fallback. With max_pixels=10 on a 32x32 file, the pre-fallback _check_dimensions(width, height, samples, max_pixels) at gpu.py:559 raises before the sparse-tile branch fires. The test passes for the wrong reason. The kwarg-recorder test directly above (test_sparse_tile_fallback_forwards_all_kwargs) already pins max_pixels plumbing via assertion on seen[0]['max_pixels']. Either drop this test as redundant, or convert it to a case where the file's full IFD fits under max_pixels but a synthetic sub-decoder failure forces the fallback to re-run with the same kwarg (closer to the recorder pattern in test_planar2_fallback_forwards_all_kwargs).

• The hand-crafted TIFF fixture helpers _write_rotated_tiled_tiff and _write_sparse_rotated_tiled_tiff are nearly identical; the sparse variant differs only in the offset/byte_count packing for the last tile. Consider factoring the shared IFD-build logic into a helper or parameterising one writer on a sparse: bool flag. About 80 lines of fixture code collapses to about 10. Not blocking; the duplication is contained and well-commented.

Nits (optional improvements)

• gpu.py:801: the comment "now-already-sliced fallback buffer" reads slightly awkwardly. "the already-sliced fallback buffer" or "the buffer (now sliced by _read_to_array)" is cleaner.

• test_gpu_fallback_forwards_kwargs_2238.py:392-393: import order. _gpu_decode is imported inside the test body. Pre-existing style in this repo (see test_predictor2_big_endian_gpu_1517.py) puts function-scope imports at the top of the test function. The current order works; flagged only for consistency.

What looks good

• The fix is small and mirrors the existing arr_was_cpu_decoded skip pattern already used by _apply_orientation_gpu (gpu.py:817-820). The new skip on _gpu_apply_window_band (gpu.py:864-873) and the multi-band shape guard (gpu.py:801-811) reuses the same flag, so future readers can see all three skips together.
• The post-decode shape check is correctly narrowed to non-CPU-decoded paths so it still catches pure-GPU kernel regressions while not falsely tripping on the now-windowed fallback output.
• The xfail flip on test_cupy_rotated_read_drops_crs (was strict=True, raises=NotImplementedError) provides automatic regression coverage; if the kwarg-forwarding breaks again, that test will raise in CI before the new test file even runs.
• Test coverage hits all four call sites: stripped (via the existing #2122 test), sparse (real round-trip on a hand-crafted rotated tiled file), planar=2 (monkeypatched _gpu_decode_single_band_tiles returning None), and decode-failure (both gpu_decode_tiles_from_file and gpu_decode_tiles patched to raise).
• The _make_kwarg_recorder helper is the right abstraction here. Asserting on seen[0] rather than chasing data shapes keeps each fallback site's contract pinned at the API boundary, independent of downstream slicing fixes.
• No backend dispatch changes; the numpy / dask / dask+cupy paths are correctly identified as unaffected (the dask+cupy chunked path already forwards kwargs via _read_geotiff_gpu_chunked).

Checklist

• Algorithm correctness: no algorithm change; kwarg-plumbing fix only
• All implemented backends produce consistent results: verified by the cross-fallback recorder tests and the 1189-test sweep
• NaN handling: unchanged; the sparse-tile fallback still goes through _finalize_eager_read for the mask
• Edge cases: window/band/max_pixels/allow_rotated each covered for at least one fallback site
• Dask chunk boundaries: N/A (eager GPU path only)
• No premature materialization or unnecessary copies: N/A
• Benchmark: N/A (bug fix)
• README feature matrix: N/A (no backend change)
• Docstrings: comments at each new branch reference issue #2238; no public API change

[brendancol]

PR Review (round 2): post-follow-up

Blockers (must fix before merge)

• None.

Suggestions (should fix, not blocking)

• None remaining from the previous review.

Nits (optional improvements)

• None remaining from the previous review.

What changed since the last review

• test_sparse_tile_fallback_respects_caller_max_pixels was dropped. The recorder test (test_sparse_tile_fallback_forwards_all_kwargs) covers max_pixels plumbing.
• _write_sparse_rotated_tiled_tiff was folded into _write_rotated_tiled_tiff via a sparse: bool kwarg. The shared IFD-build logic now lives in one place; the sparse path only adds offset=0/bytecount=0 entries for the last tile.

Disposition of round-1 findings

• Suggestion 1 (misleading max_pixels test): fixed by deletion. The duplicate-coverage rationale in the review was the right call.
• Suggestion 2 (TIFF writer duplication): fixed via the sparse=True parameter on the unified writer. Diff dropped ~93 lines of fixture code.
• Nit 1 (comment wording): dismissed. The phrase the reviewer flagged was in the round-1 review body, not the source. The actual gpu.py:804 comment ("the buffer is already sliced and will not equal the full IFD shape") reads cleanly.
• Nit 2 (import order): dismissed. Surrounding test files (e.g. test_predictor2_big_endian_gpu_1517.py) put function-scope imports inside the test body, matching the current style. The suggested move would diverge from the repo's existing convention.

Verification

• pytest xrspatial/geotiff/tests/test_gpu_fallback_forwards_kwargs_2238.py xrspatial/geotiff/tests/test_allow_rotated_no_crs_2122.py: 13 passed