rasterize: reject zig-zag duplicate-coord patterns in _check_uniform_axis

[brendancol]

Summary

• Closes a validation gap in _check_uniform_axis. The previous check
  compared abs(np.diff(coords)) against abs(expected_step), so a
  zig-zag like x = [0.5, 1.5, 0.5, 1.5] (abs-diffs all 1.0) passed
  as "uniform" and produced output with duplicate coords that broke
  .sel.
• Now compares signed diffs against the signed first interval.
  Ascending and descending axes still pass; zig-zag and duplicate-coord
  patterns raise a ValueError naming the offending axis.

Backend coverage

Validation lives in the shared _extract_grid_from_like path, so all
four backends (numpy, cupy, dask+numpy, dask+cupy) pick this up.

Test plan

• New xrspatial/tests/test_rasterize_signed_step_2566.py: ascending
  and descending uniform pass; zig-zag x and y rejected; strictly
  increasing non-uniform still rejected (the rasterize(like=...) reuses irregular coords with uniform-grid rasterizer #2168 case); single-cell
  axis short-circuits; symptom-level check that the rasterize call
  raises before it can return a DataArray with duplicate coords.
• Existing TestLikeUniformGridValidation (7 tests) still passes.
• Full test_rasterize.py: 215 passed, 2 skipped.

Closes #2566

[brendancol]

PR Review: rasterize signed-diff uniformity check (#2585)

Blockers

None.

Suggestions

• xrspatial/rasterize.py:2925-2929 -- the caller-side comment still reads "Compare abs(diff) against the first interval so the check stays agnostic to axis direction". After this PR the validator compares signed diffs against the signed first interval, so the comment now describes the old behaviour. Update it.

Nits

• xrspatial/rasterize.py:2810 -- expected_step is now only used for the == 0 and isfinite short-circuits; the actual comparison computes coords[1] - coords[0] internally. The caller still passes abs(x[1] - x[0]), which works for both gates because only magnitude matters there. Worth a follow-up: either drop the parameter or rename it so the abs-magnitude role is explicit.
• xrspatial/tests/test_rasterize_signed_step_2566.py -- the file docstring lists six pinned cases but the class has seven test methods (the trailing "no duplicate output coords" symptom check). Update the count or drop the test (it overlaps with test_zigzag_x_rejected).

What looks good

• Minimal surgical fix: signed diff against signed first interval, no other behaviour changes.
• Zero-step and non-finite short-circuits still work because the caller's abs(...) agrees with the new internal signed_step on both edge values.
• The comment at rasterize.py:2944 ("rejected non-monotonic or duplicate-valued y coords") is now accurate -- it was aspirational before.
• Tests pin ascending uniform, descending uniform, zig-zag on x and y, strictly-increasing non-uniform (#2168), and the single-cell short-circuit.
• Validation lives upstream of all four backend paths, so numpy, cupy, dask+numpy, and dask+cupy all pick this up without per-backend changes.

Checklist

• Algorithm matches issue description
• All backends produce consistent results (shared validation)
• NaN handling unchanged (non-finite short-circuit preserved)
• Edge cases covered (descending, single-cell, zig-zag)
• Dask chunk boundaries -- N/A
• No premature materialization or unnecessary copies
• Benchmark -- N/A
• README feature matrix -- N/A
• Docstrings present and accurate

[brendancol]

Follow-up review pass

Addressed in 43fb6ff8:

• Suggestion (stale caller comment at rasterize.py:2925-2929): fixed. The comment now describes the signed-diff comparison.
• Nit (test docstring listed 6 cases for 7 methods): fixed. Added the seventh entry covering the symptom-level zig-zag check.

Deferred to follow-up:

• Nit (expected_step parameter no longer drives the main comparison): not changed in this PR. The parameter is still consulted for the expected_step == 0 and np.isfinite(expected_step) short-circuits, so it is not dead. Renaming or dropping it expands the surgical scope of the bug fix without changing observable behaviour; better as its own cleanup.

Test status after follow-up: all 14 in TestSignedStepValidation_2566 + TestLikeUniformGridValidation pass.