Add native CUDA kernel for hydraulic erosion (#961)

[brendancol]

Closes #961

Summary

• Adds _erode_gpu_kernel (@cuda.jit), one CUDA thread per particle, with cuda.atomic.add for heightmap conflicts
• Replaces the CPU-fallback path for CuPy and Dask+CuPy arrays with on-device execution
• Refactors erode() to use ArrayTypeFunctionMapping dispatch instead of manual type checks
• Adds test_erosion.py (15 tests across numpy, dask+numpy, cupy, dask+cupy)
• Adds user guide notebook 21_Hydraulic_Erosion.ipynb
• Updates README feature matrix from fallback to native for GPU columns

Notes

• GPU results are non-deterministic because cuda.atomic.add ordering depends on thread scheduling. GPU tests use correlation checks instead of exact comparison.
• Float64 arithmetic in the GPU kernel to match the CPU path.
• Erosion is a global operation (particles traverse the full grid), so dask arrays are still materialized before processing. They just run on-device now instead of round-tripping through NumPy.

Test plan

• 15 new tests in test_erosion.py, all passing
• Verify GPU tests pass on CI with CUDA
• Run user guide notebook end-to-end to check plots render