Keep template backend and variance name in kriging singular-matrix fallback

[brendancol]

Closes #3288

• The K_inv is None fallback in kriging() now builds its all-NaN result on the template's backend: map_blocks + np.full_like for dask templates (stays lazy, keeps chunks and the cupy meta on dask+cupy) and np.full_like otherwise (dispatches to cupy for GPU templates). Previously the fallback always returned an eager numpy array, and for a large chunked dask template it materialised the whole grid in host memory.
• With return_variance=True the fallback's variance raster is now named <name>_variance, matching the normal path. It used to inherit the prediction's name.

Backend coverage: numpy, cupy, dask+numpy, dask+cupy. All four verified locally on a GPU host.

Test plan:

• 4 new tests in TestKriging force the fallback via monkeypatch and check name, backend type, chunks, dtype, attrs, and NaN content per backend
• pytest xrspatial/tests/test_interpolation.py: 70 passed (GPU tests included)

[brendancol]

PR Review: Keep template backend and variance name in kriging singular-matrix fallback

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• xrspatial/interpolate/_kriging.py (new map_blocks call in the K_inv is None branch): the lambda gives the dask task a lambda-<hash> name instead of an xrspatial.kriging-style name per the #3255 convention. Every other map_blocks call in this module has the same gap, and #3256 already tracks the module-wide naming sweep, so fixing only this one call here would leave the module inconsistent either way. Fine to leave for #3256.

What looks good

• The fix dispatches on the actual backend: map_blocks + np.full_like keeps dask templates lazy with their original chunks, and np.full_like dispatches to cupy for GPU templates. Verified in the new tests, including the dask+cupy meta check (result.data._meta is cupy.ndarray).
• The fallback was eager before this change even for chunked dask templates; now it never materialises the grid, which matches the memory-guard reasoning that skips the grid-sized term for dask (_check_kriging_memory(is_dask=True)).
• variance now carries <name>_variance like the normal path (line 523), and the test asserts it on all four backends.
• The monkeypatch in the tests targets the module global _build_kriging_matrix, which kriging() resolves at call time, so the forcing mechanism is sound.
• Coords, dims, attrs, and dtype assertions are explicit in test_singular_fallback_metadata rather than inferred.

Checklist

• Algorithm matches reference/paper (n/a, metadata-only fix)
• All implemented backends produce consistent results
• NaN handling is correct (all-NaN fallback by design)
• Edge cases are covered by tests (fallback forced per backend)
• Dask chunk boundaries handled correctly (no overlap needed; chunks preserved)
• No premature materialization (fallback is now lazy on dask; was eager before)
• Benchmark exists or is not needed (degenerate path, not performance-relevant)
• README feature matrix updated (n/a, no API or backend-support change)
• Docstrings present and accurate (no public signature change)