aspect: test degenerate shapes and geodesic boundary modes

[brendancol]

Closes #2742

Adds the two missing test paths the deep-sweep test-coverage pass found in the aspect module. Test-only; no source changes.

• Degenerate rasters (1x1, Nx1, 1xN) now have regression coverage for aspect, northness, and eastness. Below size 3 along a dimension the 3x3 kernel has no interior cell, so the correct output is all-NaN with the input shape preserved.
• Geodesic boundary modes nearest, reflect, and wrap are now exercised. Before this, only boundary='nan' ran through the geodesic path, leaving its lat/lon padding code untested.

Backend coverage: both gaps are tested across numpy, cupy, dask+numpy, and dask+cupy. The cupy and dask+cupy cases were run and passed on a CUDA host (73 new cases, none skipped locally).

Test plan:

• pytest xrspatial/tests/test_aspect.py xrspatial/tests/test_geodesic_aspect.py — 155 passed
• new degenerate-shape cases pass on all four backends
• new geodesic boundary cases pass on all four backends
• cupy / dask+cupy cases observed running on GPU, not skipped

[brendancol]

PR Review: aspect test coverage (degenerate shapes and geodesic boundary modes)

Test-only PR adding two missing coverage paths to the aspect module. No source changes.

Blockers

None.

Suggestions

One note for the record: in the degenerate-shape tests, the data[0, :] = 2.0 tweak only runs when shape[0] > 1, so the (1, 5) and (1, 3) cases feed a flat all-ones raster. The all-NaN assertion still holds there, but for the no-interior-cell reason rather than a flat-surface reason. That is the intended geometric edge case, so it is fine as written.

Nits

• The new _to_numpy helper in test_aspect.py duplicates the backend-unwrap logic that general_checks.py already does inline. Not worth extracting for three call sites, but it could fold into a shared helper later.

What looks good

• Both gaps are covered across all four backends (numpy, cupy, dask+numpy, dask+cupy) using the existing create_test_raster helper and the standard skip decorators.
• The cupy and dask+cupy cases ran on a CUDA host and passed, not just added behind a skip.
• Geodesic boundary tests assert the actual contract: non-default modes fill the edge ring, nan keeps it NaN, plus numpy-vs-dask/cupy parity.
• Dask chunks are clamped to the raster size for the degenerate shapes, so no chunk-larger-than-array warnings.

Checklist

• Algorithm matches reference: n/a (test-only)
• All implemented backends consistent: yes, parity asserted
• NaN handling correct: yes, that is what the tests check
• Edge cases covered: yes, that is the point of this PR
• Dask chunk boundaries handled: yes, chunks clamped
• No premature materialization: compute only at assertion time
• Benchmark: not needed (test-only)
• README matrix: not needed (no new function)
• Docstrings: n/a