Merge sink_d8 labels across dask tile boundaries (#1394)

[brendancol]

Fixes #1394.

Summary

• _run_dask_numpy in xrspatial/hydro/sink_d8.py ran per-tile CCL with globally unique IDs but never merged equivalent labels across tile boundaries, so a single connected sink that spanned a chunk showed up as several separate sinks.
• Added a union-find pass that walks every interior tile edge (4-connected) plus both diagonals (so 8-connectivity is preserved across corner-shared tiles), records label equivalences, and remaps labels to their roots via a second map_blocks pass.
• The dask+cupy path delegates to the numpy dask path, so it inherits the fix.

Test plan

• pytest xrspatial/hydro/tests/test_sink_d8.py — 40 passing (21 original + 19 new)
• pytest xrspatial/hydro/tests/ — full hydro suite still passes (772 tests)
• Original reproducer from sink_d8 dask backend splits connected sinks across tile boundaries #1394 now matches the numpy result
• Added regression tests for horizontal, vertical, diagonal, four-tile-block, and separation cases at multiple chunk shapes
• Added a _label_count_matches_numpy test so the number of unique sink labels stays equal across backends

Notes

• Cross-tile merging needs a global view of the labeled raster, so the implementation calls .compute() once on the per-tile result. The streaming benefit of dask is preserved during the per-tile CCL phase; only the boundary scan and label remap require the materialized array. CCL is fundamentally a global operation, so this matches what xrspatial/sieve.py already does for its dask path.
• Labels in the dask output are not byte-identical to the numpy output (the dask path uses position-based IDs from each tile, then merges, while the numpy path uses one position-based ID across the whole raster). The new tests check label-partition equivalence (cells in the same numpy component are in the same dask component) rather than literal ID equality.