Fix bump OOM: int32 coords, count cap, per-chunk dask partitioning (#1206)

[brendancol]

Summary

Fixes #1206. Three issues in bump() that cause wrong results and OOM on large rasters:

• uint16 coordinate overflow: locs used dtype=np.uint16, so coordinates > 65,535 wrapped silently. Now uses int32.
• Unbounded default count: count defaulted to w * h // 10 with no ceiling. Now capped at 10M bumps, plus a memory guard that raises MemoryError when the arrays would exceed 50% of available RAM.
• Closure serialization in dask paths: _bump_dask_numpy and _bump_dask_cupy captured the full locs and heights arrays in every chunk closure. At 2048x2048 default count with 128x128 chunks, that was 5 MB/chunk across 256 chunks = 1.29 GB of graph payload. Now pre-partitions bumps by chunk using dask.delayed + da.block, so each task only serializes its own subset. Graph serialization reduced ~250x.

Test plan

• Existing bump tests pass (all 10)
• New test: int32 coordinates work at 70,000 (no uint16 wrap)
• New test: default count capped for 20,000 x 20,000 raster
• New test: MemoryError raised for huge explicit count
• New test: dask graph size stays proportional to total bumps, not bumps * chunks
• New test: single-chunk dask matches numpy bitwise