Fixes #878: extend KDTree path to allocation/direction to prevent OOM

[brendancol]

Summary

• Extends the memory-safe KDTree path (previously PROXIMITY-only) to handle ALLOCATION and DIRECTION modes, preventing OOM when max_distance=inf on large dask-backed rasters
• cKDTree.query() returns both distances and nearest-target indices, which is enough to derive allocation values and compass direction angles
• Adds a memory guard in the non-KDTree fallback path (GREAT_CIRCLE or missing scipy) that raises MemoryError with actionable guidance instead of silently OOMing

Key changes

• _vectorized_calc_direction() — new array-based compass direction function matching the scalar _calc_direction() floating-point behaviour
• _stream_target_counts() — now also caches target pixel values alongside coordinates
• _collect_region_targets() — returns (coords, vals) tuple
• _kdtree_chunk_fn() / _tiled_chunk_query() — extended with mode dispatch (PROXIMITY/ALLOCATION/DIRECTION)
• KDTree gate widened — removed process_mode == PROXIMITY restriction
• Memory guard — MemoryError raised when single-chunk rechunk would exceed 80% available memory

Test plan

• test_allocation_dask_kdtree_matches_numpy (EUCLIDEAN + MANHATTAN) — consistency check against proximity
• test_direction_dask_kdtree_matches_numpy (EUCLIDEAN + MANHATTAN) — consistency check against allocation
• test_allocation_dask_kdtree_with_target_values — target filtering through KDTree
• test_direction_dask_kdtree_with_target_values — target filtering through KDTree
• test_allocation_dask_kdtree_no_targets / test_direction_dask_kdtree_no_targets — all-NaN
• test_allocation_dask_kdtree_max_distance / test_direction_dask_kdtree_max_distance — truncation
• test_great_circle_dask_unbounded_memory_guard — MemoryError for GREAT_CIRCLE
• test_no_scipy_dask_unbounded_memory_guard — MemoryError when scipy missing
• All 51 existing + new tests pass, no regressions