generate_terrain: map dask backends over an empty skeleton instead of the template

[brendancol]

Summary

• The dask backends of generate_terrain regenerate every cell from coordinates and never read the template's values. They now map the per-chunk generator over a da.empty_like(data) skeleton instead of the template array, so a NaN-filled template (e.g. from_template()'s da.full(shape, nan)) drops out of the graph instead of being memset and discarded.
• Removes the leftover data * 0 in the dask+cupy path, which built a NaN array the per-chunk cupy.zeros immediately discarded.

Output values are unchanged; this only removes wasted allocation and fill from the dask graph. Verified that the from_template -> generate_terrain graph no longer carries a full/NaN layer (only empty_like + xrspatial.terrain remain).

Closes #3574

Backends

• numpy: unchanged (eager path allocates its own output)
• cupy: unchanged
• dask+numpy: skeleton swap
• dask+cupy: skeleton swap + data * 0 removed

Test plan

• test_terrain_dask_does_not_materialize_template (worley on and off): a poisoned template whose blocks raise if computed must still produce finite terrain
• test_terrain_dask_skeleton_matches_numpy: skeleton output matches the numpy reference
• test_terrain_dask_cupy_does_not_materialize_template
• Full test_terrain.py suite green (56 passed, all four backends on the CUDA box)

[brendancol]

PR Review: generate_terrain dask backends map over an empty skeleton

Blockers

None.

Suggestions

• The new tests cover worley on and off for the dask+numpy non-materialization case, but test_terrain_dask_cupy_does_not_materialize_template only runs with worley off. The worley pre-pass map_blocks was also switched to the skeleton, so a worley-on cupy case would exercise that second call site on GPU too. Low priority, since the numpy parametrize already covers both sites and the existing test_terrain_all_nan_template_dask_cupy_matches_numpy covers the default path. (xrspatial/tests/test_terrain.py)

Nits

• The skeleton still produces one uninitialized np.empty/cupy.empty allocation per chunk feeding map_blocks. That is the expected pattern and there is no memset, so the real win is dropping the da.full(nan) fill, not the allocation. No code change needed.

What looks good

• Verified the from_template -> generate_terrain graph drops the full/NaN layer (only empty_like + xrspatial.terrain remain) and the output stays finite and matches numpy.
• Removes the leftover data * 0 in the dask+cupy path, the same NaN*0 pattern #3525 fixed on the eager backends.
• The poison-template tests are a direct proof that the template is never read: blocks that raise on compute, yet the result computes fine.
• empty_like preserves shape, dtype, and chunking, so block_info coordinates are unchanged and output values are identical (parity test passes).

Checklist

• No algorithm change (output verified identical)
• All four backends produce consistent results
• NaN handling correct (template NaN dropped, output finite)
• Edge cases covered (poison template, worley on/off)
• Dask chunk boundaries handled correctly (empty_like preserves chunking)
• Removes a premature materialization rather than adding one
• Benchmark exists (benchmarks/benchmarks/terrain.py)
• README feature matrix update not needed (no new function or backend)
• Comments added; no public API change

[brendancol]

Follow-up review (after 8a9b1eb)

The earlier Suggestion is addressed: test_terrain_dask_cupy_does_not_materialize_template is now parametrized over worley_blend [0.0, 0.2], so the worley pre-pass map_blocks call site is exercised on the dask+cupy path as well.

• Blockers: none
• Suggestions: none outstanding
• Nits: the earlier allocation note stands as documentation only; no code change warranted.

Local run: test_terrain.py template/materialize/skeleton tests pass (10 passed), full file green earlier (56 passed) on the CUDA box. Nothing new introduced by the follow-up.