Fix proximity dask fallback mutating the caller's input chunks

[brendancol]

Closes #2847

_process_dask() did raster.data = raster.data.rechunk(...) on the unbounded GREAT_CIRCLE and no-scipy fallback paths, which rebound .data on the caller's input DataArray. Code that reused the same DataArray afterward saw changed chunking.

Changes

• Rechunk into a local variable in _process_dask instead of reassigning raster.data, so the input is left untouched.
• Add regression tests asserting the input .data keeps its identity and chunking after proximity/allocation/direction on both unbounded fallback paths.

Backend coverage

The bug and fix are dask-only. Tests run on dask+numpy and dask+cupy. numpy and cupy paths never reassigned .data and are unaffected.

Test plan

• New mutation regression tests fail on the old code and pass on the fix
• Full test_proximity.py suite passes (291 passed)

[brendancol]

PR Review: Fix proximity dask fallback mutating the caller's input chunks

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• test_proximity.py: on a host without CUDA, the dask+cupy parametrization runs the same code as dask+numpy, because _backend_raster only swaps the buffer to cupy when CUDA is present. So the dask+cupy cases give real coverage only on a GPU runner. That matches the rest of the file and needs no change; just flagging that the dask+cupy assertions are duplicate work off-GPU.

What looks good

• The fix is minimal and correct. data = raster.data followed by data = data.rechunk(...) leaves the caller's DataArray untouched, since dask .rechunk returns a new array rather than mutating in place. The map_overlap call now reads from the local data.
• The dask+cupy unbounded path was already safe (it rebinds raster to a .copy() before reaching _process_dask); this fix covers the dask+numpy unbounded path and the converted dask+cupy case without touching that.
• Tests assert both object identity (raster.data is data_before) and chunking equality, and cover all three public entry points (proximity, allocation, direction) on both unbounded fallback paths. The tests fail on the pre-fix code and pass after.

Checklist

• Algorithm unchanged; only the chunk-handling side effect was removed.
• NaN handling unchanged.
• Dask chunk boundaries handled correctly; rechunk semantics preserved.
• No premature materialization or extra copies introduced.
• No README or docstring change needed (pure bug fix, no API change).