geotiff tests: consolidate input-validation cluster

[brendancol]

Closes #2430. Cluster 6 of the long-tail GeoTIFF test consolidation epic (#2424).

Summary

Fold nine top-level files into a new xrspatial/geotiff/tests/unit/test_input_validation.py:

• test_degenerate_pixel_size_2214.py
• test_geotiff_band_bool_rejection_1786.py
• test_geotiff_band_type_rejection_1910.py
• test_pixel_array_count_cap_1901.py
• test_size_param_validation_1752.py
• test_strip_zero_dims_2053.py
• test_tile_size_multiple_of_16_1767.py
• test_validate_3d_non_band_trailing_dim_2240.py
• test_window_out_of_bounds_1634.py

The new file is organised by validation axis:

1. band type / bool rejection -- band must be a non-negative int; bool / np.bool_ raise ValueError, float / str raise TypeError, across read_to_array, open_geotiff, read_geotiff_dask, read_geotiff_gpu, and read_vrt.
2. Size-parameter validation -- tile_size and read_geotiff_dask chunks must be positive, and tile_size must be a multiple of 16 when tiled=True.
3. Source-dimension validation -- zero / negative ImageWidth / ImageLength / SamplesPerPixel rejected on stripped and tiled read paths, local and HTTP, plus pixel-array count caps against IFD geometry.
4. 3D writer-dim validation -- (y, x, <non-band>) DataArray inputs rejected rather than silently written band-last.
5. Window-bounds validation -- an out-of-bounds window raises a clear ValueError on the eager and dask paths.
6. Degenerate pixel-size fail-closed -- a 1xN / Nx1 write with no transform and no opt-in raises rather than borrowing the other axis's pixel size.

One intentional dedup: the bool file's test_read_to_array_band_one_still_works is the same test as the type file's test_read_to_array_band_int_still_works (both read band 1), so only one survives. That accounts for 145 collected here vs 146 across the nine originals.

The reader.windowed release-gate row in docs/source/reference/release_gate_geotiff.rst cited test_window_out_of_bounds_1634.py; it now cites the consolidated file so the checklist-parity gate stays green.

Audit

xrspatial/geotiff/tests/CLUSTER_AUDIT_INPUT_VALIDATION.md maps every old file::test to its new file::test_id. The audit file is deleted on a final commit on this branch before merge.

Verification

• pytest xrspatial/geotiff/tests/unit/test_input_validation.py -v: 145 passed.
• pytest xrspatial/geotiff/tests/ -q: 5814 passed, 68 skipped, 2 xfailed.
• find xrspatial/geotiff/tests -name 'test_*.py' | wc -l: 236 (was 244; -9 deleted +1 added = -8).

Test plan

• New unit/test_input_validation.py passes on this checkout.
• Full xrspatial/geotiff/tests/ suite passes.
• Release-gate checklist-parity test passes.

[brendancol]

PR Review: geotiff tests: consolidate input-validation cluster

Tests-only consolidation. Nine top-level input-validation files fold into unit/test_input_validation.py, grouped into 22 classes across six validation axes. No xrspatial/geotiff source changes. I read all nine originals and the new file, ran the suite, and checked parity.

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• A couple of TestBandTypeRejection cases are really bool-rejection or positive-path checks (test_read_to_array_band_bool_still_rejected, test_read_to_array_band_zero_still_works), not type-rejection. This follows how the originals were split (the type file owned the positive still-works cases), so it is defensible. A reader skimming by class name might expect those elsewhere. Not worth the churn to move them.

What looks good

• The 145-vs-146 count delta is intentional and documented: test_read_to_array_band_one_still_works is the same test as test_read_to_array_band_int_still_works (both read band 1), so one drops. Called out in the PR body and the audit file.
• The behavior-bearing helpers carry over unchanged: the IFD byte builders (_build_classic_tiff, _short_bytes, _long_bytes), the tag-patch helpers (_find_ifd_entry_offset, _patch_inline_long), the _StaticBytesHTTPSource stub, and the MAX_PIXEL_ARRAY_COUNT monkeypatch with its try/finally restore.
• The per-file _gpu_only markers are replaced with the shared requires_gpu from _helpers/markers.py, same cupy.cuda.is_available() semantics.
• The release-gate checklist row that cited the deleted test_window_out_of_bounds_1634.py now points at the consolidated file, so the checklist-parity gate stays green. Easy to miss.
• isort and flake8 both pass (lines under 100).

Checklist

• N/A algorithm vs reference (tests-only)
• All four read backends still covered (numpy eager, dask, gpu-gated, vrt)
• NaN handling unchanged (no logic touched)
• Edge cases preserved (zero/negative dims, out-of-bounds windows, degenerate 1xN/Nx1)
• N/A dask chunk boundaries
• N/A premature materialization
• N/A benchmark
• N/A README feature matrix (no new public API)
• Docstrings: module-level and per-class present