Document NaN propagation in edge_detection docstrings

[brendancol]

Closes #3676

• Adds a Notes section to the sobel_x, sobel_y, prewitt_x, prewitt_y, and laplacian docstrings stating that NaN input cells propagate to their 3x3 neighborhood, and that the default boundary='nan' leaves the outer one-cell ring NaN.
• Tightens test_nan_in_input_propagates to assert the exact documented footprint: the 3x3 block around the NaN cell is NaN and nothing else is.
• Also commits the error-handling sweep's state CSV row for this module.

Behavior is unchanged; this documents what the code already does. The footprint is identical on numpy, cupy, dask+numpy, and dask+cupy (checked on a GPU host).

Test plan:

• pytest xrspatial/tests/test_edge_detection.py - 87 passed, including the cupy and dask+cupy classes

[brendancol]

PR Review: Document NaN propagation in edge_detection docstrings

Blockers (must fix before merge)

• none

Suggestions (should fix, not blocking)

• none

Nits (optional improvements)

• xrspatial/edge_detection.py (Notes sections, all five functions): "every output cell whose 3x3 neighborhood contains a NaN becomes NaN" is exact for interior NaNs, but under boundary='wrap' a NaN on one edge also poisons cells on the opposite edge (checked: NaN at (2,0) on a 5x6 raster with boundary='wrap' makes rows 1-3 of the last column NaN, since the wrapped neighborhood crosses the seam). Wording like "3x3 neighborhood (as extended by the boundary mode)" would cover that case.

What looks good

• The documented claim was verified against observed behavior, not just read from the code: one interior NaN produces exactly a 3x3 NaN block, identical on numpy, cupy, dask+numpy, and dask+cupy.
• test_nan_in_input_propagates now asserts the full footprint (NaN block and everything else non-NaN) instead of a single cell, so the docstring claim is pinned by a test.
• Behavior is unchanged; the diff is docstrings, one test, and the sweep's state CSV row, which belongs in this PR per the sweep contract.

Checklist

• Algorithm matches reference/paper (n/a, docs-only)
• All implemented backends produce consistent results (footprint checked on all four)
• NaN handling is correct (and now documented)
• Edge cases are covered by tests
• Dask chunk boundaries handled correctly (n/a, no code change)
• No premature materialization or unnecessary copies (n/a)
• Benchmark exists or is not needed (docs-only)
• README feature matrix updated (n/a, no new functions or backend changes)
• Docstrings present and accurate (see nit on wrap wording)

[brendancol]

PR Review: follow-up after 907f8ed

The single nit from the first pass is addressed: all five Notes sections now read "3x3 neighborhood (as extended by the boundary mode)", which covers the wrap-seam case observed during review. Re-ran pytest xrspatial/tests/test_edge_detection.py: 87 passed.

No blockers, no suggestions, no nits remaining.