Add diffusion test coverage for cupy/dask+cupy, reflect boundary, and edge shapes

[brendancol]

Closes #3422

diffuse() dispatches to all four backends through ArrayTypeFunctionMapping, but the test file only exercised numpy and dask+numpy. This adds the missing coverage. Tests only, no source changes.

• cupy and dask+cupy parity against the numpy result
• reflect boundary mode across numpy, dask, cupy, and dask+cupy (previously only nearest and wrap were tested)
• 1x1 single-pixel and Nx1 / 1xN strip rasters, including a chunked dask strip
• all-NaN and Inf inputs

Backend coverage: numpy, cupy, dask+numpy, dask+cupy. The four GPU tests ran on a CUDA host before this PR; CI reruns them on its own GPU runner. The reflect/strip/NaN/Inf cases also confirmed all paths produce correct results today, so #3422 is a coverage gap, not a latent bug.

Test plan:

• pytest xrspatial/tests/test_diffusion.py -> 39 passed (was 24)
• 14 new tests pass, including 4 cupy / dask+cupy tests executed on a CUDA host

[brendancol]

PR Review: Add diffusion test coverage for cupy/dask+cupy, reflect boundary, and edge shapes

Test-only PR. No source changed, so the correctness checks apply to the tests themselves, not to new library code.

Blockers

None.

Suggestions

None.

Nits

• test_inf_input_propagates_like_nan: the assertion on the Inf cell's neighbour is np.isnan(out[1, 2]) or np.isinf(out[1, 2]). That is a deliberately loose smoke check given Inf arithmetic is brittle, which is fine, but it would still pass if the behaviour flipped from inf to nan. Worth a one-line comment saying that is intentional so a future reader does not tighten it by mistake.

What looks good

• All four backends are exercised. The cupy and dask+cupy tests are guarded with cuda_and_cupy_available, so they skip on non-GPU hosts and run on CI's GPU runner.
• Cross-backend parity uses the shared general_checks.py helpers with reasonable tolerances (1e-7 cupy, 1e-6 dask+cupy for reduction-order differences).
• The edge cases line up with the issue: 1x1, Nx1, 1xN, all-NaN, Inf, plus the previously untested reflect boundary.
• The dask strip test chunks across the strip (chunks=(4,1)), so it exercises the overlap path instead of collapsing to one chunk.

Checklist

• All implemented backends produce consistent results (parity tests)
• NaN handling correct (equal_nan, all-NaN and propagation tests)
• Edge cases covered
• Dask chunk boundaries handled (strip chunked across the overlap)
• No source changes -> no benchmark/README/docstring updates needed
• Tests pass locally: 39 passed, including 4 GPU tests on a CUDA host