Fixes #880: replace single-chunk rechunk in cost_distance with iterative tile Dijkstra (#888)

cost_distance() had two OOM paths when max_cost=inf (the default) or the
implied overlap exceeded chunk dimensions: (1) rechunking both arrays into
a single chunk, materialising the full raster, and (2) allocating a
full-size heap inside the kernel.

Replace the single-chunk fallback with a streaming boundary-only iterative
Dijkstra that processes tiles one at a time, storing only 1-pixel-wide
boundary strips between iterations — O(sqrt(N)) memory instead of O(N).

Algorithm: Gauss-Seidel forward/backward sweeps seed each tile's Dijkstra
from converged neighbour boundaries (cardinal + diagonal along shared
edges + corner seeds for 8-connectivity), iterate until no boundary
changes, then lazily assemble the final result via da.map_blocks.

Author: brendancol