Add GPU and edge-case test coverage for spline()

[brendancol]

What

Test-only PR from a scoped test-coverage sweep on the spline interpolator. No source changes.

spline() registers four backends through ArrayTypeFunctionMapping, but TestSpline only exercised numpy and dask+numpy. The cupy and dask+cupy paths, including the _tps_cuda_kernel CUDA kernel, had no test coverage. Kriging already has GPU parity tests; spline did not.

Changes

• test_cupy_matches_numpy and test_dask_cupy_matches_numpy on TestSpline, guarded with the existing cuda_and_cupy_available / dask_array_available decorators.
• test_two_point_affine_fit covering the n == 2 least-squares affine branch in _tps_build_and_solve.
• test_output_metadata asserting the output keeps the template's coords, dims, attrs, and name.

Backend coverage

numpy, cupy, dask+numpy, dask+cupy all exercised for spline now.

Test plan

• pytest xrspatial/tests/test_interpolation.py::TestSpline -- 9 passed on a CUDA host (both GPU parity tests ran, not skipped)
• pytest xrspatial/tests/test_interpolation.py -- 45 passed

No GH issue is linked: issue creation was blocked by the local auto-mode classifier during this run.

[brendancol]

PR Review: Add GPU and edge-case test coverage for spline()

Test-only PR. No source changes, so the checks below apply to the tests, not to new production code.

Blockers

None.

Suggestions

None.

Nits

• test_cupy_matches_numpy and test_dask_cupy_matches_numpy both use a 3x3 square grid. A non-square grid (e.g. 2x3) would exercise the kernel's row/col indexing harder, but the existing kriging GPU tests in this file use square grids too, so matching the convention is fine.

What looks good

• Both GPU tests reuse the existing cuda_and_cupy_available / dask_array_available decorators and the _to_numpy helper, matching the kriging tests in the same file.
• test_two_point_affine_fit pins the n==2 affine-fit branch with a hand-checkable oracle: (0,0)->10 and (2,0)->20 give a pure x-gradient, so x=1 reads 15 and x=2 reads 20.
• test_output_metadata checks coords, dims, attrs, and name. It goes one step past the IDW metadata test by also asserting attrs equality.
• The dask_cupy test uses chunks=(2,2) on a 3x3 grid, so it covers ragged (2,1) chunk boundaries.

Checklist

• All four spline backends covered now (numpy, cupy, dask+numpy, dask+cupy): yes
• GPU tests guarded with skip decorators: yes
• Edge cases (n==2 branch, metadata) covered: yes
• Reference values hand-checked, not just "runs without crashing": yes
• Temp files uniquely named: n/a, none
• Docstrings on tests: present