Add GPU (CuPy) backend for cost_distance

[brendancol]

Summary

• Replaces CPU spill-to-host fallback with native CUDA iterative parallel relaxation (parallel Bellman-Ford)
• Each CUDA thread processes one pixel per iteration, checking all 4/8 neighbours for shorter paths; the wavefront advances at least one pixel per iteration
• Convergence in O(height + width) iterations with early termination when no pixel changes
• CuPy path: seeds sources at cost 0, iterates relaxation kernel, converts inf/over-budget to NaN
• Dask+CuPy: bounded max_cost uses da.map_overlap with per-chunk GPU relaxation (out-of-core, no OOM); unbounded falls back to CPU iterative tile Dijkstra
• Tests extended with cupy and dask+cupy across 8 parametrized test functions + 2 new GPU-specific tests
• README updated: Cost Distance row now shows checkmarks for CuPy and Dask GPU columns

Fixes #905

Test plan

• All 44 cost_distance tests pass (including 16 new cupy/dask+cupy variants)
• Verified on GPU (NVIDIA RTX A6000) with cupy 13.6.0
• Uniform friction, analytic grids, barriers, multiple sources, max_cost truncation, target_values, 4/8 connectivity all validated against numpy reference
• Bounded dask+cupy hits map_overlap GPU path (verified no iterative warning)
• Unbounded dask+cupy correctly falls back to CPU iterative tile Dijkstra
• Output types verified: cupy returns cupy array, dask+cupy returns dask+cupy chunks