test(convolution): backend, boundary, and metadata coverage

[brendancol]

What

Adds tests for xrspatial/convolution.py. Before this, the test file had 5 tests
that only touched numpy error paths and the default boundary mode, so branch
coverage sat at 52%.

convolve_2d and convolution_2d dispatch across numpy, cupy, dask+numpy, and
dask+cupy, but only the numpy path was exercised. The boundary parameter has four
public modes and only the default was tested. convolution_2d re-attaches
coords/dims/attrs/name and nothing checked that it does.

New tests

• Backend parity: convolution_2d numpy vs dask+numpy, cupy, and dask+cupy.
• Boundary modes: nan/nearest/reflect/wrap on numpy+dask, and a cupy/dask+cupy
  vs numpy comparison over all four modes. Plus the invalid-boundary ValueError.
• Metadata: convolution_2d preserves coords, dims, attrs, and name; custom name.
• NaN inside the raster propagates through the kernel window.
• Degenerate shapes (1x1, 1xN, Nx1) return all-NaN without crashing.
• Previously untested public functions: calc_cellsize (res attr and km->m unit
  conversion), annulus_kernel, custom_kernel valid/non-ndarray/even-dims, and
  circle_kernel bad radius/unit.

Verification

Ran on a CUDA host, so the cupy and dask+cupy tests executed for real, not just
static review. Full file: 26 passed, including the 3 GPU tests. flake8 and isort
clean.

No source change. No bug surfaced by the new tests. This is test-only.

Branch coverage went from 52% to 72% as measured on CPU (NUMBA_DISABLE_JIT=1).
The cupy kernel lines can't be seen by CPU coverage instrumentation, but they run
under the 3 passing GPU tests.

Found by the test-coverage deep sweep.

[brendancol]

PR Review: test(convolution): backend, boundary, and metadata coverage

Test-only PR. xrspatial/tests/test_convolution.py gains 189 lines closing the
backend-parity, boundary-mode, and metadata gaps in convolution.py, plus first
tests for calc_cellsize, annulus_kernel, custom_kernel, and circle_kernel.
convolution.py is untouched. I read the source and the general_checks.py
helpers, ran the full file, and confirmed the assertions line up.

Verification (ran locally on the CUDA host)

• pytest xrspatial/tests/test_convolution.py: 26 passed.
• The three GPU tests (test_convolution_2d_numpy_equals_cupy,
  test_convolution_2d_numpy_equals_dask_cupy,
  test_convolution_2d_boundary_modes_gpu_match_numpy) ran for real, not skipped.
• flake8 and isort clean.
• Import resolved into the PR worktree, not the main editable install.

Blockers

None.

Suggestions

None blocking.

Nits

• test_convolution_2d_boundary_modes_gpu_match_numpy re-implements the
  boundary-mode loop by hand (.data.get() / .data.compute().get()) instead of
  going through assert_boundary_mode_correctness. That is because the helper
  only compares numpy against a dask arg, so there is no shared path for the
  cupy/dask+cupy comparison. Fine as written; worth a helper only if a later PR
  adds more GPU boundary tests.
• The degenerate-shape and NaN-interior tests pin current behavior (all-NaN
  output, NaN propagation through the plus-kernel window) rather than a reference
  value. That matches the algorithm and the PR's intent, so no change needed.

What looks good

• Backend parity reuses the assert_numpy_equals_* helpers, so the numpy path is
  the reference for all three other backends.
• Boundary coverage hits all four public modes (nan/nearest/reflect/wrap)
  on numpy, dask, cupy, and dask+cupy, plus the invalid-mode ValueError.
• The NaN-interior test picks cells whose 3x3 circle window overlaps the seeded
  NaN and one that does not, so it checks propagation and non-propagation.
• Kernel-generator error paths (non-ndarray, even dims, bad radius/unit) are
  covered.

Checklist

• Algorithm matches reference (test-only; assertions match source behavior)
• All implemented backends produce consistent results (parity tests, GPU ran)
• NaN handling is correct (interior propagation + boundary='nan' edges)
• Edge cases covered (1x1/1xN/Nx1 degenerate, invalid boundary, bad kernel)
• Dask chunk boundaries handled (boundary-mode helper compares dask vs numpy)
• No premature materialization or unnecessary copies (test-only)
• Benchmark exists or is not needed (test-only, no perf change)
• README feature matrix updated (n/a, no new public function)
• Docstrings present and accurate (n/a for tests; source docstrings unchanged)