geotiff tests: consolidate rotated-CRS cluster

[brendancol]

Summary

Folds three rotated / dropped-CRS test files into a single parametrized file at xrspatial/geotiff/tests/read/test_crs.py under the layout from epic #2390 PR 1.

Files folded:

• test_allow_rotated_geotiff_2115.py
• test_allow_rotated_crs_drop_2126.py
• test_allow_rotated_no_crs_2122.py

The HTTP rotated test (test_http_dask_allow_rotated_2130.py) is left for PR 9 (integration).

What changed

• New read/__init__.py and read/test_crs.py with 22 parametrized tests over (scenario, chunks) where scenario is rotated_no_crs / rotated_with_crs / axis_aligned_with_crs and chunks is eager / dask.
• GPU-eager and dask+CuPy variants kept under @requires_gpu.
• VRT eager + chunked rotated cases parametrized over the same chunks axis.
• Cross-file consumers of _write_rotated_tiff (test_georef_status_2136.py, test_lazy_finalization_parity_2162.py) now import from the new module.
• Release-gate and reference docs that cite the old filenames point at read/test_crs.py.

Audit

CLUSTER_AUDIT_PR3.md maps every old file::test to its new file::test_id. Two HTTP cases from 2115 were dropped; the same coverage already exists in test_http_dask_allow_rotated_2130.py (PR 9 cluster). The audit file is deleted on a follow-up commit before approval.

Test plan

• pytest xrspatial/geotiff/tests/read/ -v: 22 passed
• pytest xrspatial/geotiff/tests/ -x -q: 5705 passed, 68 skipped
• File count: 3 deleted, 1 added (net -2)

Closes #2396. Part of epic #2390.

[brendancol]

PR Review: geotiff tests: consolidate rotated-CRS cluster

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• xrspatial/geotiff/tests/read/test_crs.py (the _build_* callers around lines 357 / 374 / 390 / 410): temp filenames are built with f"..._{chunks}.tif"; when chunks is None this writes literal _None.tif paths under pytest's tmp_path. Functional but a bit ugly. Worth swapping for the test ID label ("eager" / "dask") via a small lookup.
• xrspatial/geotiff/tests/read/test_crs.py: _build_rotated_no_crs returns (src, arr) but the default-raises tests discard arr. Trivial; src, _ = ... would make the intent obvious. Same shape in _build_axis_aligned_with_crs.
• xrspatial/geotiff/tests/CLUSTER_AUDIT_PR3.md: the audit table lumps the two dropped HTTP cases as "covered by PR 9". Correct, but naming the specific tests in test_http_dask_allow_rotated_2130.py (e.g. test_http_dask_rotated_default_raises, test_http_dask_rotated_allow_rotated_reads) would make the trail explicit so a future reviewer does not have to open the integration file to confirm.

What looks good

• Parametrization is the right shape: (scenario, chunks) with descriptive IDs (rotated_no_crs, rotated_with_crs, axis_aligned_with_crs crossed with eager / dask), and GPU variants stay behind @requires_gpu without bleeding the import gate into the rest of the file.
• Both rotated TIFF writers are preserved as distinct test paths rather than collapsed: hand-rolled with the full _GEO_KEYS_4326 block, and the tifffile-based one with a Geographic-only GeoKey. Different GeoKey-parser branches; worth keeping both.
• Unit tests for _extract_transform and _populate_attrs_from_geo_info are parametrized over the attrs they assert on (crs / crs_wkt / transform), which is cleaner than the original three-separate-functions layout.
• Cross-file consumers of _write_rotated_tiff (test_georef_status_2136.py, test_lazy_finalization_parity_2162.py) are updated to import from the new module, with a relocation comment citing the epic PR. The two doc paths (release_gate_geotiff.rst, geotiff.rst) that cited the old filenames also point at the new location.
• 5705 passing on the full geotiff suite locally; file count moves -2 as the epic intended.

Checklist

• Algorithm matches reference/paper (N/A: tests-only restructure)
• All implemented backends produce consistent results (eager / dask / GPU eager / dask+CuPy covered)
• NaN handling is correct (N/A)
• Edge cases are covered by tests (axis-aligned regression guard preserved)
• Dask chunk boundaries handled correctly (chunks=2 exercised across all rotated cases)
• No premature materialization or unnecessary copies (N/A)
• Benchmark exists or is not needed (N/A)
• README feature matrix updated (N/A)
• Docstrings present and accurate (module docstring plus per-test docstrings preserved or improved)
• CLUSTER_AUDIT_PR3.md present; will be removed in follow-up commit

[brendancol]

Follow-up review

All three nits from the prior review are addressed in 89d0078:

• Filename labels: new _chunks_label helper maps None to eager and 2 to dask, so tmp_path now contains ..._eager.tif / ..._dask.tif instead of ..._None.tif. Applied across all five parametrized cases that previously interpolated the raw chunks value (the no-CRS default-raises, no-CRS optin-reads, rotated-with-CRS hand-rolled, rotated-with-CRS geokey-only, GPU, and the VRT pair).
• Unused-binding cleanup: src, _ = _build_*(...) in the default-raises and axis-aligned tests so it is obvious the array fixture is intentionally unused at those sites.
• Audit specifics: CLUSTER_AUDIT_PR3.md now names the two replacing PR 9 tests (test_http_dask_rotated_default_raises, test_http_dask_rotated_allow_rotated_reads) on the dropped HTTP rows.

The audit file was then removed in 283b600 before approval, per the epic plan. The 22 read/ tests still pass locally.

No new findings on the follow-up pass.