Add edge_detection tests for Inf, boundary NaN, degenerate shapes, and multi-chunk dask

[brendancol]

Closes #3682

Test-only change from the test-coverage audit of edge_detection. The source is untouched; every new test pins behavior that was verified correct on a CUDA host but had no coverage.

• Inf and -Inf input: exact expected 3x3 neighborhood per operator (inf where the kernel weight has one sign, NaN where inf meets -inf or a zero weight), unaffected cells stay 0.
• NaN at the raster corner under all four boundary modes, with numpy vs dask+numpy (multi-chunk), cupy, and dask+cupy parity. Earlier NaN tests only used an interior cell on numpy; NaN-input backend divergence has shipped before (kde, kde/line_density NaN inputs: backends disagree and the auto extent collapses to zeros #3628).
• Degenerate shapes: 1x1 (all-NaN with boundary='nan', 0 with 'reflect'), 1x6 and 6x1 strips, (0, 5) empty raster.
• Multi-chunk dask equivalence (chunks smaller than the array) for dask+numpy and dask+cupy; the existing dask tests all used one chunk, so chunk stitching was never exercised from this file.
• boundary='invalid' raises ValueError (same test the other convolution-backed modules have).
• lat/lon dim names pass through.

Backends: numpy, cupy, dask+numpy, dask+cupy all exercised; the GPU tests ran locally, not just in CI.

Test plan:

• pytest xrspatial/tests/test_edge_detection.py on a CUDA host: 222 passed (was 87)
• flake8 on the test file: no new errors (line 22 E302 pre-exists, tracked by Style: flake8 E201 in edge_detection kernel literals #3675)
• branch coverage of xrspatial.edge_detection stays 100%

Also carries the sweep state CSV row for this audit.

[brendancol]

PR Review: Add edge_detection tests for Inf, boundary NaN, degenerate shapes, and multi-chunk dask

Test-only PR, source untouched. Ran the new suite locally on a CUDA host: 222 passed, GPU tests executed rather than skipped.

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• Integer dtype input is still untested. #3682 calls out input-space gaps, and convolve_2d promotes int32 to float32; a small test pinning the promoted dtype and a known value would close that corner (new sections of xrspatial/tests/test_edge_detection.py).
• TestInfHandling.test_negative_inf (test_edge_detection.py, ~line 396) only asserts an inf-or-nan mask. The -inf case is the +inf expected block negated with NaN kept, so it could pin exact values the way test_inf_neighborhood does.

Nits (optional improvements)

• The module-level inf = np.inf / nan = np.nan aliases (test_edge_detection.py, ~line 358) leak into the module namespace for the sake of one parametrize table. Inlining np.inf/np.nan in the table would keep the namespace clean, at the cost of longer lines.

What looks good

• test_inf_neighborhood pins exact per-operator 3x3 expectations instead of a loose isfinite check, including the 0-weight -> NaN cells.
• NaN-at-corner parity runs all 4 boundary modes on all 4 backends; this is the combination that has caught real divergence elsewhere (kde #3628).
• Multi-chunk dask tests (chunks=(4,5) on 9x11, chunks=(3,4) on 8x10) finally exercise chunk stitching from this file; previous dask tests were all single-chunk.
• test_invalid_boundary_raises matches the pattern aspect/morphology/diffusion/bilateral already use.
• The pre-existing E302 at line 22 was left alone (tracked by #3675), no drive-by edits.

Checklist

• Algorithm matches reference/paper (n/a, test-only; expectations match probed behavior)
• All implemented backends produce consistent results
• NaN handling is correct
• Edge cases are covered by tests
• Dask chunk boundaries handled correctly
• No premature materialization or unnecessary copies
• Benchmark exists or is not needed (missing, tracked separately in #3672)
• README feature matrix updated (n/a)
• Docstrings present and accurate (n/a; docstring gaps tracked in #3674/#3676)

[brendancol]

PR Review: follow-up after 65687d4

Re-checked the branch after the review fixes.

Blockers

None.

Suggestions

None. Both earlier suggestions are addressed:

• TestIntegerDtype.test_int32_promotes_to_float32 pins the float32 promotion and compares against the float64-input result for all five operators.
• test_negative_inf now asserts the exact negated neighborhood via the shared INF_NEIGHBORHOODS table instead of an inf-or-nan mask.

The nit is also done: the module-level inf/nan aliases are gone, with the expected blocks built from np.inf/np.nan directly.

Verification

• 227 tests pass locally on a CUDA host (was 222 at first review, 87 before the PR).
• flake8 on the file reports only the pre-existing E302 at line 22 (#3675).

Nothing further from me.