Fix hypsometric_integral dask+cupy backend (#2525)

[brendancol]

Closes #2525.

Summary

• _hi_dask_cupy previously converted cupy chunks to numpy and then called _hi_dask_numpy, returning a dask+numpy result for dask+cupy input. Downstream ArrayTypeFunctionMapping dispatch silently routed through the numpy path.
• After this change, the painted output is re-wrapped through map_blocks(cupy.asarray, ...) so each chunk lands back as a cupy array. The dask graph stays lazy and is_dask_cupy(result) is True for dask+cupy inputs.
• The numpy reduce in the middle is kept as-is. The per-zone HI lookup is a Python dict so the reduce must pass through host memory; only the painted output needs to match the input backend.

Backend coverage

• numpy: unchanged, returns numpy ndarray
• cupy: unchanged, returns cupy ndarray
• dask+numpy: unchanged, returns dask Array with numpy chunks
• dask+cupy: now returns dask Array with cupy chunks (was: numpy chunks)

Test plan

• New parametrized regression test test_hi_preserves_backend_2525 asserts the output backend matches the input across all four backends.
• Existing test_hypsometric_integral.py suite still passes (33 tests).
• Existing test_zonal.py suite still passes (125 tests).
• Verified the new test fails with the fix reverted (confirms it catches the regression).

Audit trail

Surfaced by /sweep-metadata on the zonal module, 2026-05-27. Cat 5 (backend-inconsistent metadata), severity MEDIUM.

[brendancol]

Review summary

The fix is correct and minimal. _hi_dask_cupy now wraps the painted output via map_blocks(cupy.asarray, ..., meta=cupy.array(())), mirroring the pattern in _apply_dask_cupy (zonal.py:1720, 1731). All four backends remain wired in ArrayTypeFunctionMapping, and the new parametrized test verifies both DataArray-level and chunk-level backend identity.

Blockers

None.

Suggestions

None blocking. The extra map_blocks(cupy.asarray, ...) adds one host-to-device copy per chunk, but this is unavoidable given that the per-zone lookup is a Python dict and must go through host memory anyway.

Nits applied in follow-up commit

• Replaced the locally redeclared _has_cuda_and_cupy() with the shared has_cuda_and_cupy from xrspatial.utils, and swapped da is None for has_dask_array() to match the rest of the test suite.
• Dropped the em-dash in the _hi_dask_cupy docstring.

What looks good

• Test asserts chunk-level type via .blocks[0, 0].compute(), not just the DataArray wrapper.
• PR body confirms the test fails when the fix is reverted.
• Docstring updated to explain why the round-trip through host memory exists.

Pre-existing items (not from this PR)

• xrspatial/zonal.py:39 F401 not_implemented_func unused and xrspatial/zonal.py:455 E501 both exist on the merge base.