Consolidation cluster 14: GPU (4 sub-PRs) (long-tail epic #2424)

Follow-up to epic #2424 / #2390. Cluster 14 of 16. The largest remaining cluster by far.

Will be 4 PRs. Single issue for tracking; four PRs against this issue for actual landings. Each sub-PR is independently reviewable.

Scope

43 GPU-flavoured tests grouped by concern under a new gpu/ directory.

Sub-PR layout

Sub-PR A: `gpu/test_reader.py` (about 10 files)

GPU reader paths: window-band, stripped (multiband, no-georef-window, forwarding), nodata, miniswhite-band-first, chunks-out-of-core, sidecar-georef-parity, read-eager-url, declared-dtype.

Files: test_gpu_window_band_1605.py, test_gpu_stripped_multiband.py, test_gpu_stripped_no_georef_window_1753.py, test_gpu_stripped_forwarding_1732.py, test_gpu_nodata_1542.py, test_gpu_miniswhite_band_first_2097.py, test_gpu_chunks_out_of_core_1876.py, test_gpu_sidecar_georef_parity_2324.py, test_read_geotiff_gpu_url_eager_2161.py, test_chunked_gpu_declared_dtype_1909.py.

Sub-PR B: `gpu/test_writer.py` (about 8 files)

GPU writer paths: attrs, band-first, compression-modes, cpu-fallback-codecs, nan-sentinel, overview-inplace, overview-mode-and-compression-level, to_geotiff-gpu-fallback.

Files: test_gpu_writer_attrs_1563.py, test_gpu_writer_band_first_1580.py, test_gpu_writer_compression_modes_2026_05_11.py, test_gpu_writer_cpu_fallback_codecs_2026_05_12.py, test_gpu_writer_nan_sentinel_1599.py, test_gpu_writer_overview_inplace_1948.py, test_gpu_writer_overview_mode_and_compression_level_1740.py, test_to_geotiff_gpu_fallback_1674.py.

Sub-PR C: `gpu/test_codec.py` (about 12 files)

GPU codec paths: nvcomp (batched, single-alloc, decompress-cumsum-offsets, batch-upload), nvjpeg (encode-stream-sync, 2k-single-alloc), jpeg-gpu, lerc-valid-mask-gpu, predictor (2 / 3 GPU), jpeg-interop-reject.

Files: test_nvcomp_batch_compress_batched_1712.py, test_nvcomp_batch_upload_p3.py, test_nvcomp_decompress_cumsum_offsets_1950.py, test_nvcomp_from_device_bufs_single_alloc_1659.py, test_nvjpeg_encode_stream_sync_2212.py, test_nvjpeg2k_single_alloc_2107.py, test_jpeg_gpu_1549.py, test_lerc_valid_mask_gpu.py, test_predictor2_big_endian_gpu_1517.py, test_predictor3_int_dtype_gpu_1933.py, test_gpu_jpeg_interop_reject_issue_D_1845.py.

Sub-PR D: `gpu/test_kernels_and_kwargs.py` (about 13 files)

GPU kernel + kwarg surface: cuda-preflight, kwarg-rename, strict-fallback, fallback-forwards-kwargs, kvikio-batched-pread, gds (chunked-parity, fallback-batched-d2h), orientation-gpu, size-param-validation-gpu-vrt, mask-nodata-gpu-vrt, open-geotiff-on-gpu-failure, crs-fail-closed-gpu.

Files: test_gpu_cuda_preflight_1903.py, test_gpu_kwarg_rename_1560.py, test_gpu_strict_fallback_1516.py, test_gpu_fallback_forwards_kwargs_2238.py, test_kvikio_batched_pread_1688.py, test_gds_chunked_gpu_parity_1896.py, test_gds_fallback_batched_d2h_1552.py, test_orientation_gpu.py, test_size_param_validation_gpu_vrt_1776.py, test_mask_nodata_gpu_vrt_2052.py, test_open_geotiff_on_gpu_failure_1615.py, test_crs_fail_closed_gpu_1929.py.

Golden-corpus GPU files (test_golden_corpus_dask_gpu_1930.py, test_golden_corpus_gpu_1930.py) belong to the golden-corpus relocation cluster (issue 16). Hands off here.

Constraints

Tests-only.
Markers from _helpers/markers.py; make_minimal_tiff from _helpers/tiff_builders.py. Apply @requires_gpu on every test.
Descriptive IDs (e.g. id="nvcomp-batched[float32-zlib]").
A CLUSTER_AUDIT_GPU_<letter>.md per sub-PR, deleted before merge of that sub-PR.

Verification

pytest xrspatial/geotiff/tests/gpu/ -v  # requires CUDA
pytest xrspatial/geotiff/tests/ -x -q

File count drops by 39 (43 deleted, 4 added).

Note on register pressure

This project has hit GPU register-pressure issues before (see memory note in CLAUDE.md / project history). When consolidating the writer-overview and nvcomp tests, watch for kernel-launch parametrise IDs that combine many heavy paths; if the test compilation slows materially, split parametrise IDs into per-codec sub-cases.