Make dask float polygonize chunk-invariant for rtol

[brendancol]

Closes #2666.

What

• numpy CCL groups adjacent float pixels with abs(value - reference) <= atol + rtol*abs(reference), where reference is the higher-ij pixel in scan order, so the test is asymmetric in rtol. The dask cross-chunk merge compared unordered (min, max) ranges, so its merge decision flipped with chunk order.
• Each chunk-boundary polygon now carries its internal-edge pixel values keyed by global (col, row). The cross-chunk merge applies numpy's higher-ij-is-reference orientation at the exact pixel pair straddling a boundary, so a chunked float raster produces the same polygon partition as the unchunked input.
• Integer rasters keep strict equality (empty cell map, range fallback).

Backends

Float path: numpy / cupy (float routes through numpy), dask+numpy, dask+cupy. The merge change lives in the shared dask path used by both dask+numpy and dask+cupy. Integer behaviour is unchanged.

Known limitation

8-connectivity at large rtol still diverges between numpy and dask. That is a pre-existing ring-merge topology limitation (present on main before this change), not the rtol orientation bug, so it is left for a separate fix.

Test plan

• pytest xrspatial/tests/test_polygonize_issue_2666.py (29 passed)
• Full polygonize suite still green (test_polygonize.py, _issue_2583, _issue_2172, coverage files): 291 passed, 16 skipped
• Fuzz across shapes / chunkings / rtol: 4-connectivity now matches numpy in every case (was 26 failures on main)

[brendancol]

PR Review: Make dask float polygonize chunk-invariant for rtol

Blockers

None.

Suggestions

• 8-connectivity at large rtol still diverges between numpy and dask (a dense 4x4 float array with rtol=0.1 gives different polygon counts). This is pre-existing on main, not introduced here. Tracing shows it is a ring-merge topology issue, not a close-value orientation issue: it fails identically across chunk shapes that do not even split the relevant pixels, so the cross-chunk boundary logic is not the cause. The PR scopes around it correctly. Worth a follow-up issue so the limitation is tracked outside the PR body.

Nits

• _ranges_close (polygonize.py:295) and its interval-overlap fallback are now reached only for integer rasters and the defensive empty-cell-map case. For integers the check reduces to equality, so the overlap branch no longer fires in practice. Harmless and pre-existing, but a one-line comment marking it as a fallback path would help the next reader.

What looks good

• The fix targets the root cause: numpy CCL uses the higher-ij pixel as the rtol reference, and the merge now replicates that orientation at the exact boundary pixel pair instead of comparing unordered ranges.
• Boundary-cell maps are O(chunk perimeter), keyed by global coords, so the dask path gains no full-array materialization.
• Integer rasters keep strict equality through the empty cell map and range fallback.
• NaN boundary cells are skipped during extraction, so they never merge, matching numpy.
• Tests cover the repro, its reverse, column/row orientation, disconnected close-valued pixels, several chunkings, the default tolerance path, integer invariance, and dask+cupy.

Checklist

• Algorithm matches numpy CCL scan-order orientation
• Implemented backends consistent (4-conn float verified; integer unchanged)
• NaN handling correct
• Edge cases covered by tests
• Dask chunk boundaries handled correctly
• No premature materialization or unnecessary copies
• Benchmark not needed (bug fix)
• README/docs unchanged (no public API change)
• Docstrings updated for the new tuple field and helper

[brendancol]

Follow-up review (after review commits)

Re-reviewed the two follow-up commits.

Disposition of prior findings

• Suggestion (8-conn divergence at large rtol): deferred to follow-up issue #2677. The fix is a separate ring-merge topology problem, pre-existing on main, and would expand this PR well beyond the rtol orientation bug.
• Nit (_ranges_close now a fallback path): fixed. Its docstring now states it is the fallback for integers and the empty-cell defensive case, with float merging routed through _cells_close_directed.

New changes

• Added dask+cupy parity tests for the repro and its reverse, guarded by the cuda/cupy marker. They skip locally without a GPU.

No new blockers, suggestions, or nits. The 4-connectivity rtol parity is complete and the existing suite stays green (47 passed, 3 skipped on the polygonize parity files).