geotiff tests: consolidate georef and rotation cluster (#2435)

[brendancol]

Closes #2435 (cluster 11 of the long-tail GeoTIFF test consolidation epic #2424). Tests-only.

Summary

Fold twelve top-level georef / no-georef / rotation test files into a single xrspatial/geotiff/tests/read/test_georef.py, grouped by concern:

Read side:

• test_georef_edges.py (Add geotiff edge-case tests: IFD chain loops, header parsing, descending-y, non-georeferenced #1482) -- non-georef reads, south-up writes
• test_georef_resolver_parity_2211.py (Refactor GeoTIFF Phase 2: centralize transform/georef contract (PR-B of #2211) #2225) -- resolve_georef parity across reader / writer call sites
• test_georef_status_2136.py (geotiff: canonical georef_status attribute to disambiguate CRS vs transform presence #2136) -- canonical georef_status attr for the five reader states
• test_no_georef_attr_migration_2133.py (geotiff: finish migration from integer coord sentinel to explicit no-georef attr #2133) -- no-georef marker migration from coord dtype to attrs
• test_no_georef_marker_2120.py (geotiff: to_geotiff silently strips georef on int64 step-1 user coords #2120) -- int64 step-1 user grids keep their georef
• test_no_georef_windowed_coords_1710.py (open_geotiff windowed reads of non-georef TIFFs produce float64 half-pixel-shifted coords #1710) -- windowed reads emit the same integer pixel coords as full reads
• test_rotated_transform_attr_1764.py (to_geotiff silently drops rotation/shear from attrs['transform'] #1764) -- _transform_from_attr rejects rotated / sheared 6-tuples
• test_rotated_affine_attr_2129.py (geotiff: surface rotated_affine 6-tuple on the DataArray attrs #2129) -- allow_rotated=True surfaces attrs['rotated_affine']
• test_rotated_typed_error_2267.py (Rotated GeoTIFF errors bypass the typed RotatedTransformError exception #2267) -- GeoTIFF and VRT rotated reads raise the same typed error

Write side (kept in the same file under a write-side heading because they share the rotated-affine and no-georef-marker fixtures; splitting into write/test_georef.py would scatter one concern across two modules):

• test_degenerate_georef_1945.py (geotiff: 1xN / Nx1 georeferenced writes silently strip transform #1945) -- 1xN / Nx1 / 1x1 georeferenced writes preserve their transform
• test_no_georef_writer_round_trip_1949.py (geotiff: to_geotiff turns no-georef int64 pixel coords into a fake transform on round-trip #1949) -- no-georef files keep integer coords across read/write/read
• test_to_geotiff_drop_rotation_2216.py (to_geotiff silently drops rotated_affine instead of failing closed #2216) -- to_geotiff refuses to silently drop a rotated affine

GPU gating moves to the shared requires_gpu marker and gpu_available probe from _helpers/markers.py. The per-issue rotated-TIFF and VRT writers keep their distinct byte layouts and are suffixed by issue number to avoid collision. The status suite imports _write_rotated_tiff from read/test_crs.py, which owns the rotated-CRS read surface and is left untouched. No assertion changed.

Audit

CLUSTER_AUDIT_GEOREF.md maps every old file to its new section. The audit file is deleted in a final pre-merge commit.

Verification

• pytest xrspatial/geotiff/tests/read/test_georef.py -q: 171 passed (same total as the twelve source files).
• pytest xrspatial/geotiff/tests/read/ -q: 711 passed, 4 skipped.
• pytest xrspatial/geotiff/tests/ --co -q: 5890 collected, no errors.
• File count drops by 11 (12 deleted, 1 added).

Test plan

• Consolidated read/test_georef.py passes on this checkout (GPU present).
• Full geotiff test suite collects.
• CI green across numpy / cupy / dask+numpy / dask+cupy.
• Audit file deleted on the pre-merge commit.

[brendancol]

PR Review: geotiff tests: consolidate georef and rotation cluster (#2435)

Tests-only consolidation. Twelve top-level files fold into read/test_georef.py. No source under xrspatial/geotiff/ changed, so the algorithm / backend-dispatch / performance sections of the checklist do not apply. I verified the move preserved every test rather than re-deriving correctness.

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• read/test_georef.py:81 imports _write_rotated_tiff as _write_rotated_tiff_crs with a # noqa: E402, but the import sits in the normal top-of-file block, not after code. flake8 with --select=E402 is clean without it, so the marker can be dropped.

What looks good

• Test count is identical: 144 unique test functions in both the old twelve files and the new module, 171 collected items each. comm shows zero names dropped.
• Every assertion is carried over verbatim; the only changes are mechanical (helper renames, shared marker).
• Helper-name collisions across the source files (_write_rotated_tiff, _write_rotated_vrt, _make_da, _rotated_dataarray, _ROTATED_TUPLE) are resolved by suffixing each with its issue number, so the distinct byte layouts stay distinct.
• The status suite keeps importing _write_rotated_tiff from read/test_crs.py with the same (path, arr) call shape; the imported helper's signature is unchanged, so the dependency holds and read/test_crs.py is untouched.
• _NO_GEOREF_KEY resolves to the same string from _coords and _geotags, so consolidating the import to _coords does not change any marker comparison.
• GPU gating moved to the shared requires_gpu marker; the one spot that read the old _HAS_GPU global now calls gpu_available(), which returns the same bool.
• flake8 is clean; full suite collects 5890 with no errors.

Checklist

• [n/a] Algorithm matches reference/paper -- no algorithm changed
• All implemented backends produce consistent results -- GPU + dask + numpy paths preserved, all pass on this checkout
• NaN handling is correct -- unchanged from source files
• Edge cases are covered by tests -- 1x1 / Nx1 / 1xN / no-georef / rotated all carried over
• Dask chunk boundaries handled correctly -- unchanged
• No premature materialization or unnecessary copies -- tests-only
• [n/a] Benchmark exists or is not needed -- tests-only
• [n/a] README feature matrix updated -- no new function
• Docstrings present and accurate -- module + per-section docstrings explain provenance

[brendancol]

Follow-up review (round 2)

The one nit from the first pass is fixed in 58bd26c: the unnecessary # noqa: E402 on the read/test_crs import is gone, and flake8 stays clean.

The audit file CLUSTER_AUDIT_GEOREF.md is deleted in 1e5c023 per epic #2424. Nothing else outstanding; the diff is now tests-only with no leftover scaffolding. 171 tests still pass on this checkout.