Fix cost_distance leaking dask task name as output .name

[brendancol]

Closes #3344

cost_distance returned a DataArray whose .name depended on the backend: numpy and cupy gave name=None, but dask+numpy and dask+cupy leaked the internal dask graph name (_trim-..., asarray-...) as the user-visible .name. That token is random per run, so .to_dataset() and any name-keyed logic became nondeterministic on the dask paths.

• Force the result name to the input raster's name with .rename(raster.name). Passing name= to the constructor isn't enough because xarray treats name=None as "infer from data" and lets the dask token through.
• Add a parametrized regression test over all four backends and both named and unnamed input.

Backends: numpy, cupy, dask+numpy, dask+cupy all verified (CUDA host).

Test plan:

• test_result_name_matches_input passes on all four backends (8 cases, no skips on this GPU host)
• Full test_cost_distance.py suite passes (63 passed)

[brendancol]

PR Review: Fix cost_distance leaking dask task name as output .name

Blockers

None.

Suggestions

None.

Nits

None.

What looks good

• The fix is minimal and correct. .rename(raster.name) is the right tool: I confirmed that passing name=None to the xr.DataArray constructor does not override a dask array's internal name, but .rename(None) does force it, so the constructor-only approach would not have fixed the unnamed-input case.
• Only .name changes; data, coords, dims, and attrs are untouched.
• The regression test runs all four backends and both named and unnamed inputs (8 cases). On this CUDA host none of them skip, so the cupy and dask+cupy paths are genuinely exercised.
• The supports_dataset path is unaffected: a Dataset input calls the function per variable with raster.name already equal to the variable name, so the rename keeps the variable key. Verified by running a Dataset through the dask backend.
• The inline comment explains the non-obvious reason .rename is needed instead of a constructor name=, which will save the next reader a debugging session.

Checklist

• Behavior matches intent (name consistent across backends)
• All four backends produce consistent .name
• No data/coords/dims/attrs regression
• Edge cases covered (named + unnamed input, all backends)
• No premature materialization or extra copies (rename is metadata-only)
• Benchmark not needed (metadata-only fix)
• README / docs not applicable (no API change)