polygonize: add atol/rtol parameter coverage on cupy and dask+cupy backends

[brendancol]

Description

The polygonize() function exposes atol and rtol parameters for grouping near-equal float pixels into the same polygon. These tolerance values flow through every backend:

• numpy: _polygonize_numpy(values, mask, conn8, transform, atol, rtol)
• cupy: _polygonize_cupy(...) at polygonize.py:808 forwards atol/rtol to either _polygonize_numpy (float dtypes, lines 823-825 / 829-830) or to the integer GPU CCL path (where the args are documented no-ops)
• dask / dask+cupy: _polygonize_dask(...) at polygonize.py:1719 plumbs the values through dask.delayed(_polygonize_chunk)(..., atol, rtol) and into _bucket_key_for_value(boundary_by_value, val, atol, rtol) for cross-chunk merging

Existing tests cover atol/rtol end-to-end on the numpy and dask+numpy backends (test_polygonize_strict_float_equality, test_polygonize_dask_multi_chunk_strict_float, test_polygonize_custom_atol_intermediate, etc.), but no test exercises non-default atol/rtol values on the cupy or dask+cupy backends.

This is the same dispatcher-silently-drops-backend-kwarg pattern that has shipped as a defect on adjacent surfaces (see the GeoTIFF max_cloud_bytes, dtype, and mask_nodata dispatcher tests). A regression in the GPU backend dispatchers that dropped atol/rtol on its way to _polygonize_numpy would change the polygon count silently on float rasters and would not be caught by any current test.

Scope

Add pure-test coverage:

1. atol=0, rtol=0 on a float cupy / dask+cupy raster recovers strict-equality grouping and matches the numpy result on the same input.
2. An intermediate atol value (between two known float steps) picks the same split on cupy / dask+cupy as it does on numpy.
3. Integer cupy / dask+cupy rasters still ignore atol/rtol (matching the numpy / dask+numpy behaviour already pinned by test_polygonize_integer_default_unchanged_by_atol and test_polygonize_dask_integer_atol_ignored).

Source is not touched. Tests only.

Test plan

• New tests in xrspatial/tests/test_polygonize_atol_rtol_backend_coverage_2026_05_27.py
• All four backends covered by parity comparisons against numpy
• Run on a CUDA host so the cupy and dask+cupy branches execute, not just the skip-marker