Guard cost_distance numpy dask chunk path with _check_memory (#3343)

[brendancol]

Closes #3343

What

• Call _check_memory(h, w) inside the _make_chunk_func closure before running _cost_distance_kernel, so the numpy da.map_overlap path raises a clean MemoryError (pointing at max_cost= or smaller chunks) instead of an opaque allocator failure on an oversized chunk.

This was the only allocation path in the module without a guard. The numpy (_cost_distance_numpy), cupy/dask+cupy (_check_gpu_memory), and iterative tile Dijkstra paths already had one.

Backend coverage

• numpy: unchanged (already guarded).
• dask+numpy: now guarded per chunk (the fix).
• cupy / dask+cupy: unchanged (already guarded via _check_gpu_memory).

Test plan

• New test_dask_map_overlap_chunk_memory_guard_raises forces the map_overlap path with a finite max_cost and a chunk size larger than the overlap pad, mocks _available_memory_bytes low, and asserts a MemoryError mentioning max_cost at compute time.
• Confirmed the test fails on main (DID NOT RAISE) and passes with the fix.
• Full test_cost_distance.py suite passes (56 passed), covering numpy, dask+numpy, cupy, and dask+cupy.

[brendancol]

PR Review: Guard cost_distance numpy dask chunk path with _check_memory (#3343)

Blockers

None.

Suggestions

None.

Nits

None.

What looks good

• The fix is one line in the right place. _make_chunk_func's closure receives source_block, whose shape already includes the map_overlap depth, so _check_memory(h, w) measures the actual per-chunk allocation the kernel is about to make.
• Backend parity is the point of the change: numpy direct (_cost_distance_numpy), cupy/dask+cupy (_check_gpu_memory), and the iterative tile path all guarded this already. This was the lone gap.
• _check_memory is a no-op for normal chunks (it only raises above 50% of available RAM), so well-sized dask graphs are unaffected.
• The new test forces the map_overlap path on purpose (finite max_cost, chunk size larger than the overlap pad), mocks memory low, and asserts the MemoryError surfaces at compute time. I confirmed it fails on main (DID NOT RAISE) and passes with the fix, so it actually pins the bug.

Checklist

• Algorithm unchanged; only a pre-allocation guard added
• Backends consistent (numpy direct + dask numpy chunk now both guarded)
• NaN handling unaffected
• Edge case covered (oversized chunk raises a clear MemoryError)
• Dask chunk path handled (guard runs inside the task, fires at compute)
• No premature materialization or extra copies
• Benchmark not needed (no perf-path change)
• README/docs not applicable (no public API change)
• Pure bug fix, no docstring changes needed