geotiff: cover GPU writer degenerate shapes (1x1/1xN/Nx1)

[brendancol]

Closes #2984.

Adds degenerate-shape (1x1 / 1xN / Nx1) round-trip coverage for the GPU write path. The read side and the dask streaming writer already cover these shapes; the GPU writer was the remaining gap (the smallest shape elsewhere in gpu/test_writer.py was 2x2).

• test_write_geotiff_gpu_degenerate_round_trip: direct _write_geotiff_gpu, parametrized over 1x1 / 1xN / Nx1 and none / deflate compression.
• test_to_geotiff_dask_gpu_degenerate_round_trip: the dask+cupy write path via to_geotiff(gpu=True), parametrized over the three shapes.

Both supply an explicit attrs['transform'] since pixel size cannot be inferred from a single-element coord axis, matching the read-side degenerate fixture. Tests use the existing requires_gpu marker so non-CUDA hosts skip.

No source changes. I checked that the GPU and dask+gpu write paths already handle these shapes before writing the tests, so this is coverage, not a fix.

Backend coverage: cupy (direct GPU writer) and dask+cupy (dask+gpu write path). The numpy eager writer and dask+numpy streaming writer already had this coverage.

Test plan:

• 9 new tests run and pass on a CUDA host.
• On a non-GPU host they skip via requires_gpu.

This PR also normalizes the sweep state CSV to one physical line per record, fixing pre-existing union-merge newline corruption.

[brendancol]

PR Review: geotiff, cover GPU writer degenerate shapes (1x1/1xN/Nx1)

Test-only PR. No source changes, so the review is about whether the new tests are valid and have teeth, not about algorithm changes.

Blockers

None.

Suggestions

None blocking. One optional note below.

Nits

• The dask+gpu test only exercises compression="none", while the direct-writer test covers both none and deflate. Adding deflate to the dask+gpu parametrization would close the small asymmetry, but it is not required: the codec path is shared and already covered on the direct writer.

What looks good

• The gap is real and the scope is right. Read-side degenerate shapes already cover all four backends (read/test_degenerate_shapes.py) and the dask streaming writer covers them (integration/test_dask_pipeline.py); the GPU write path was the only hole.
• The explicit attrs['transform'] matches how the read-side degenerate fixture handles the single-element coord axis, so the two suites stay consistent.
• Assertions check both shape and pixel values via assert_array_equal, so a dropped window or a wrong-shape write would fail. I confirmed the shape mismatch raises.
• Temp files include the issue number and the shape/compression id, so no collisions across parametrized cases or parallel worktrees.
• Tests use the existing requires_gpu marker, so non-CUDA hosts skip rather than error.
• The state CSV normalization to one physical line per record fixes the pre-existing union-merge newline corruption.

Checklist

• Algorithm matches reference: n/a (test-only)
• All implemented backends consistent: cupy + dask+cupy write paths now covered; numpy/dask+numpy already covered
• NaN handling correct: n/a (no nodata in fixtures)
• Edge cases covered: yes, that is the point of the PR
• Dask chunk boundaries: dask+cupy path exercised on degenerate chunks
• No premature materialization: n/a
• Benchmark: not needed
• README matrix: not needed (no new public function)
• Docstrings present: yes on every new test

[brendancol]

Follow-up review

The one nit from the first pass is addressed in 357b454: the dask+gpu degenerate test is now parametrized over none and deflate, matching the direct-writer test. The degenerate suite is 12 tests now (3 shapes x 2 codecs for the direct writer, 3 shapes x 2 codecs for dask+gpu), all passing on a CUDA host.

No remaining findings. Test-only change, no source touched.