Add native Dask+CuPy backends for hydrology core functions (#952)

[brendancol]

Summary

• Replaces CPU fallback with native GPU tile kernels for six hydrology functions (flow_accumulation, watershed, basin, stream_order, stream_link, snap_pour_point) when running on Dask+CuPy arrays
• Each tile now runs existing CUDA kernels directly with seed injection at boundaries, keeping data GPU-resident through the iterative tile sweep
• Adds a native CUDA kernel for snap_pour_point's single-GPU CuPy path (previously fell back to CPU)
• Updates README feature matrix: all six functions now show native support across all four backends

What changed

Per-tile GPU kernels with seed injection: The Dask+CuPy path previously converted CuPy chunks to NumPy, ran the CPU tile kernel, then converted back. Now each tile runs the same GPU frontier-peeling kernels used by the single-GPU path, with external boundary values injected before the peeling loop starts. Seeds are transferred CPU-side only at tile boundaries (small O(edge_length) strips), while all tile-interior computation stays on GPU.

snap_pour_point native CuPy: Added _snap_pour_point_gpu CUDA kernel where each thread handles one pour point's windowed max search. The flow accumulation array stays on GPU instead of being pulled to CPU.

stream_link tile-aware kernel: Added _stream_link_find_ready_tile CUDA kernel that uses global coordinate offsets for position-based link IDs, so tile-local results are consistent with full-array results.

Test plan

• All 158 existing + new tests pass for all six modules
• New dask+cupy tests with multiple chunk sizes and random acyclic grids for each function
• Verified dask+cupy output matches dask+numpy output exactly for basin (pre-existing tile-sweep convergence issue affects both backends identically)