cost_distance: add CuPy and Dask+CuPy GPU backends

[brendancol]

Context

PR #859 added cost_distance() with NumPy and Dask+NumPy backends. GPU backends (CuPy, Dask+CuPy) were intentionally deferred because the algorithm requires a fundamentally different approach on GPU.

Problem

The current implementation uses multi-source Dijkstra with a binary min-heap, which is inherently sequential — it processes pixels in priority-queue order. This maps poorly to GPU parallelism where thousands of threads need independent work.

Possible GPU algorithms

• Parallel label-correcting (Bellman-Ford variant): Wavefront propagation where all pixels at the current frontier are relaxed in parallel. Simple to implement on GPU but may require many iterations to converge.
• Delta-stepping: Bucketed relaxation that groups edges by cost into "delta" buckets. Pixels within each bucket can be relaxed in parallel. Better work-efficiency than naive Bellman-Ford but more complex to implement.
• Parallel Dijkstra with work-efficient heap: Use a GPU-friendly priority queue (e.g., bucket queue backed by atomic operations). Harder to implement but closest to the CPU algorithm's efficiency.

References

• Meyer & Sanders, "Delta-stepping: a parallelizable shortest path algorithm" (2003)
• Ortega-Arranz et al., "A new GPU-based approach to the Shortest Path problem" (2013)
• Davidson et al., "Work-Efficient Parallel GPU Methods for Single-Source Shortest Paths" (2014)

Scope

• CuPy backend for cost_distance()
• Dask+CuPy backend for cost_distance()
• Update README feature matrix GPU columns
• Add GPU test parametrization to test_cost_distance.py