Add Dinf flow accumulation (#938)

[brendancol]

Summary

• flow_accumulation() silently returned wrong results for Dinf input (every cell = 1.0) because _code_to_offset truncated float angles to ints that never matched any D8 code
• Auto-detects D8 vs Dinf from input values, then branches to the appropriate algorithm
• Dinf path uses Tarboton's proportional flow splitting between two downslope neighbors
• All four backends: numpy, cupy, dask+numpy, dask+cupy

What changed

xrspatial/flow_accumulation.py

• _classify_flow_block / _detect_flow_type: scan input values to distinguish D8 codes from Dinf angles (handles the 0.0 ambiguity between D8 pit and Dinf east-angle)
• _angle_to_neighbors: decompose a Dinf angle into two downstream neighbors with proportional weights
• _flow_accum_dinf_cpu: Kahn's BFS with fractional splitting
• GPU kernels (_init_accum_indegree_dinf, _pull_from_frontier_dinf, _flow_accum_dinf_cupy): frontier-peeling with inlined angle decomposition, reuses _find_ready_and_finalize from D8
• Dask tile sweep (_flow_accum_dinf_tile_kernel, _compute_seeds_dinf, etc.): boundary-propagation with fractional cross-tile seeds
• flow_accumulation() docstring updated, Tarboton (1997) reference added

xrspatial/tests/test_flow_accumulation.py

• Detection tests (D8 codes, Dinf angles, all-NaN, -1.0 pit)
• Correctness tests (cardinal east chain, pit center, 50/50 split, cascading chain, NaN handling)
• End-to-end: flow_direction_dinf(elevation) -> flow_accumulation()
• Cross-backend: dask with 6 chunk configs, cupy, dask+cupy

Test plan

• All 45 tests pass (pytest xrspatial/tests/test_flow_accumulation.py)
• 24 existing D8 tests unchanged and passing
• Smoke test: flow_direction_dinf -> flow_accumulation on random 20x20 elevation, all values >= 1.0