Reset dask graph-token .name on pathfinding outputs (#3652)

[brendancol]

Closes #3652

On dask and dask+cupy backends, a_star_search and multi_stop_search returned DataArrays named after the internal dask graph token ('concatenate-', 'array-') while numpy/cupy returned unnamed results. xr.DataArray picks up data.name when no name is passed, and dask arrays expose their task key there.

• Reset .name = None post-construction in both functions (same fix as zonal zonal.apply() output .name is inconsistent across backends #2611 and focal focal_stats and hotspots: result .name differs across backends #2733; slope has the same reset).
• Add test_output_name_consistent_across_backends, parametrized over numpy / dask+numpy / cupy / dask+cupy (GPU cases skip when CUDA is absent).

All four backends verified locally (CUDA available on this box).

Test plan:

• pytest xrspatial/tests/test_pathfinding.py (61 passed)
• probe script confirms .name is None on all four backends for both functions

https://claude.ai/code/session_0155N4QGamQVxgpAAPbpQNq4

[brendancol]

PR Review: Reset dask graph-token .name on pathfinding outputs (#3652)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• xrspatial/tests/test_pathfinding.py:1150 — _name_check_backends() overlaps with the module-level _backends list defined at line 273. Reasonable as-is, since _backends deliberately excludes the GPU backends (those tests use the @cuda_and_cupy_available decorator instead), but a shared "all runnable backends" helper in general_checks.py would serve future cross-backend metadata tests too.

What looks good

• Both public functions have exactly one DataArray construction site, so the post-construction reset covers every backend path (numpy, cupy, dask, dask+cupy, bounded, HPA*).
• Post-construction assignment is the correct mechanism here: passing name=None to the constructor does not help because xarray falls back to the dask array's .name attribute. Same approach as slope, zonal (#2611), focal (#2733), and proximity (#2728).
• @supports_dataset on multi_stop_search re-keys per-variable results in a dict, so the None name cannot leak into Dataset variable names.
• The regression test asserts the actual property (.name is None) for both functions and parametrizes over all four backends with runtime gating, so it degrades to numpy+dask on CPU-only hosts instead of skipping wholesale.

Checklist

• Algorithm matches reference: n/a (metadata-only change)
• All implemented backends produce consistent results: yes, verified by the new test on a CUDA host
• NaN handling: untouched
• Edge cases covered: single construction site per function, so no uncovered paths
• Dask chunk boundaries: untouched
• No premature materialization: attribute assignment only
• Benchmark: not needed
• README feature matrix: no API change
• Docstrings: no signature change; existing docstrings make no claims about .name