Add golden-value tests pinning edge_detection to scipy.ndimage

[brendancol]

Closes #3677

• Adds TestGoldenScipyParity to xrspatial/tests/test_edge_detection.py: a 5x6 integer input with hardcoded expected outputs for sobel_x, sobel_y, prewitt_x, prewitt_y, and laplacian, generated with scipy.ndimage 1.16.1 (default mode='reflect')
• Asserts the boundary='reflect' full arrays match the reference exactly, and the default boundary='nan' interiors match the reference interiors
• Includes the reference-validation sweep state row for edge_detection (verdict: MATCHES)

Test-only change, no source edits. The golden tests run on the numpy backend; the existing tests in this file already check cupy, dask+numpy, and dask+cupy against numpy, so the pin covers all four backends transitively.

Background: I validated all five functions against scipy.ndimage on a Gaussian hill, a tilted plane with known analytic gradient, and a seeded random-walk surface. Interior and edge pixels matched for every boundary mode, and the tilted plane matched the analytic response exactly. The comparison script is in #3677. Nothing diverged; this PR just makes sure the test suite would notice if it ever does.

Test plan:

• pytest xrspatial/tests/test_edge_detection.py - 97 passed (87 existing + 10 new)
• flake8 on the test file: no new findings (one pre-existing E302 left alone)

[brendancol]

PR Review: Add golden-value tests pinning edge_detection to scipy.ndimage

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• TestGoldenScipyParity pins boundary='reflect' (full array) and the default boundary='nan' (interior). nearest and wrap stay covered only by assert_boundary_mode_correctness, which checks numpy against dask rather than against a reference. That seems like the right scope call: the reflect full-array already pins the stencil at every pixel plus one edge convention, and per-mode edge handling belongs to the convolution module. Flagging it so the choice is on record, not asking for more arrays.

What looks good

• Expected arrays were generated with scipy.ndimage 1.16.1 and the class docstring says exactly which calls and mode produced them, so the values can be regenerated and checked by hand (xrspatial/tests/test_edge_detection.py:127).
• assert_array_equal (exact equality) is justified here and the comment explains why: the input is integer-valued, and xrspatial matched ndi.correlate bit for bit during validation.
• The boundary='nan' interior test reuses the same golden interiors, so a regression in default-mode interior math gets caught without depending on edge conventions.
• Test-only diff plus the sweep state row; no source changes, matching what issue #3677 scoped.

Checklist

• Algorithm matches reference (validated against scipy.ndimage 1.16.1; script in #3677)
• All implemented backends produce consistent results (existing parity tests in this file cover cupy/dask+numpy/dask+cupy; golden pin on numpy transfers through them)
• NaN handling is correct (nan-boundary interior asserted; NaN propagation already tested elsewhere in the file)
• Edge cases are covered by tests (edges pinned under reflect; nan edges asserted by existing tests)
• Dask chunk boundaries handled correctly (unchanged by this PR; existing dask tests pass)
• No premature materialization or unnecessary copies (test-only)
• Benchmark exists or is not needed (not needed for tests)
• README feature matrix updated (not applicable, no new functions)
• Docstrings present and accurate (test class docstring records reference tool and version)

97 tests pass locally including the 10 new ones; flake8 adds no new findings (the one E302 predates this PR).