Add proximity coverage tests: int rasters, bounded GPU metrics, empty input (#3139)

[brendancol]

Closes #3139

Test-only PR; xrspatial/proximity.py is untouched.

• Integer-dtype rasters (int32) now run on all four backends, bounded and unbounded, pinned against the float64 numpy baseline (24 tests + 4 explicit-target_values tests). The bounded dask case is the interesting one: boundary=np.nan on an int array casts the pad to INT_MIN, a finite non-zero "target" value, and the only thing neutralizing those phantom targets is that the coordinate-grid pads are real NaNs.
• The bounded dask+cupy path (_process_dask_cupy) was only ever exercised with EUCLIDEAN. Added MANHATTAN and GREAT_CIRCLE x {proximity, allocation, direction} against the numpy baseline, with a spy asserting the bounded GPU path actually ran (6 tests). Mutation check: zeroing the halo depth flips all 6 red.
• Empty (0, N) / (N, 0) rasters now have a fail-fast pin: they raise (currently IndexError from the coordinate-endpoint lookup) instead of returning garbage (6 tests).

Backend coverage: numpy / cupy / dask+numpy / dask+cupy all executed locally on a CUDA host.

Test plan:

• 40 new tests pass on a CUDA host
• full test_proximity.py passes (450 passed, 0 skipped)
• mutation check on the bounded-GPU halo flips the new tests red; source restored byte-identical (md5)

[brendancol]

PR Review: Add proximity coverage tests: int rasters, bounded GPU metrics, empty input (#3139)

Blockers

None.

Suggestions

• Int coverage was int32-only. uint8 is the usual land-cover dtype and exercises a different NaN-pad cast (lands on a non-target 0 instead of int32's phantom-target INT_MIN). Fixed in be721dd by parametrizing dtype.
• test_integer_raster_explicit_target_values ran proximity only; allocation and direction never saw integer pixels with explicit target_values. Fixed in be721dd by parametrizing func.

Nits

• The empty-raster test is numpy-only. That's fine: the raise happens in _process before backend dispatch (the max_possible_distance endpoint lookup), so one backend covers all four. No change needed.

What looks good

• The spy on _process_dask_cupy means the bounded-GPU metric tests can't silently pass through a fallback path if dispatch thresholds drift.
• The mutation check (zeroing the halo depth) flipped all 6 bounded tests red, and the source was restored byte-identical.
• The int tests hit both the unbounded dask KDTree path (max_distance=inf) and the bounded map_overlap halo path (max_distance=10), which is where the NaN-pad cast lives.

Checklist

• Algorithm reference: n/a, test-only
• Backend consistency: 72 new tests across numpy / cupy / dask+numpy / dask+cupy, executed on a CUDA host
• NaN handling: pinned via the int-pad tests
• Edge cases: empty-raster raise pinned
• Dask chunk boundaries: bounded halo path now exercised with non-default metrics
• Benchmark / README / docstrings: not applicable, test-only