test(proximity): cover degenerate shapes, non-default metrics, attrs

[brendancol]

Closes #2692. Test-only; no source changes.

Adds 65 tests to test_proximity.py. All run and pass on a CUDA host across numpy, cupy, dask+numpy, and dask+cupy. None of these pins a bug; the current behavior is correct everywhere. They lock the behavior so a regression shows up in CI.

What's covered:

• Degenerate raster shapes (1x1 single pixel, Nx1 column strip, 1xN row strip) for proximity, allocation, and direction on all four backends. These shapes are where the line-sweep kernel boundaries (_process_proximity_line) and the GPU brute-force kernel grid sizing (_proximity_cuda_kernel) tend to break. Expected values are hand-checked; a mutation of one pinned direction expectation confirms the tests have teeth.
• MANHATTAN and GREAT_CIRCLE on allocation and direction across all four backends. Before this, only proximity exercised non-default metrics cross-backend. Both metrics match the numpy baseline.
• Attrs preservation with non-empty res/crs attrs (the shared test_raster fixture carries no attrs, so the existing attrs check compares two empty dicts). Includes a bounded-dask test that exercises the res-attr-driven padding path, since proximity reads attrs['res'] via get_dataarray_resolution to size map_overlap depth.

Full test_proximity.py suite: 134 passed (was 69).

Note for a future accuracy pass: a res attr that disagrees with the actual coordinate spacing mis-sizes the bounded-dask map_overlap depth (raises or produces wrong boundary results). That's a source fragility, out of scope for this test-only PR.