aspect: add independent rectangular-cell oracle for planar aspect

[brendancol]

Closes #2768

What this does

The planar aspect tests only prove the four backends agree with each other. They share the same _cpu/_gpu kernels, so a defect common to every backend slips through. The missing-cellsize bug (#2760) is exactly that: planar aspect divides the Horn gradients by 8 but never by cellsize_x/cellsize_y, so it is wrong on non-square cells while every backend still matches. The QGIS fixture is a real external reference but uses square cells, so it never touches rectangular cells or checks that cell size comes from the coordinates.

This adds an analytic oracle for planar aspect:

• Builds a plane z = gx*X + gy*Y on a coordinate grid with no res attr, so resolution must be recovered from the x/y coordinate spacing.
• Asserts against the closed-form Horn compass value, not against another backend.
• Square-cell oracle passes unconditionally.
• Rectangular-cell oracle is xfail(strict=False) referencing Planar aspect() ignores cell size; wrong for non-square cells #2760. It flips to XPASS automatically once that fix lands, which flags the marker for removal.

This PR is test-only. It does not touch aspect.py; the source fix is #2760's scope.

Backend coverage

numpy, dask+numpy, cupy, dask+cupy. The cupy/dask+cupy oracle variants are gated behind the existing CUDA+CuPy skip marker.

Test plan

• pytest xrspatial/tests/test_aspect.py -- 165 passed, 16 xfailed (numpy + dask; cupy paths skipped, no GPU here)
• Square-cell oracle passes on numpy and dask
• Rectangular-cell oracle xfails against current main as expected (confirmed it reproduces the Planar aspect() ignores cell size; wrong for non-square cells #2760 buggy value 275.71 vs oracle 315.0 for xres=10, yres=1)
• flake8 clean on the added lines

When #2760 merges, the rectangular oracle should start XPASSing; remove the xfail markers at that point.

[brendancol]

PR Review: aspect: add independent rectangular-cell oracle for planar aspect

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• _aspect_oracle hand-copies the compass conversion from aspect._cpu (lines 74-81). That duplication is the whole point of an independent oracle, but it means the two can drift apart if the source convention ever changes. The square-cell tests re-check the oracle against the real function on every run, so this is fine as written. A one-line comment saying the conversion is duplicated on purpose would save the next reader some confusion. (xrspatial/tests/test_aspect.py:265)
• The flat case gx == gy == 0 has a branch in _aspect_oracle but no test feeds it, so the -1 path is dead code in this PR. Either drop the branch or add a (0.0, 0.0) parametrize case. A flat-surface oracle case is cheap and worth having. (xrspatial/tests/test_aspect.py:271)

Nits (optional improvements)

• In the cupy oracle tests, the inner if backend == 'dask+cupy' and not has_dask_array(): pytest.skip(...) only matters on a machine with CuPy but no dask, which is rare. Harmless, just slightly defensive. (xrspatial/tests/test_aspect.py:325)

What looks good

• The oracle asserts against an analytic value, not another backend. That is the gap the issue called out.
• Rectangular cells with no res attr force resolution to come from coordinate spacing, which is the exact path the missing-cellsize bug breaks.
• The xfail(strict=False) pointing at #2760 is the right call: green CI now, automatic XPASS when the fix lands.
• Reuses _to_numpy and the existing skip markers; coverage spans numpy, dask, cupy, dask+cupy.

Checklist

• Oracle matches the Horn compass convention in aspect._cpu
• Square-cell oracle passes on all available backends
• Rectangular-cell oracle xfails pending #2760
• Resolution is taken from coordinates (no res attr)
• Test-only change; aspect.py untouched
• flake8 clean on added lines
• Flat-surface (-1) oracle case not exercised

[brendancol]

Follow-up review

The two suggestions from the first pass are addressed in 12be968:

• The flat-surface case (0.0, 0.0) is now in the square oracle params, so the -1 branch in _aspect_oracle is exercised. xrspatial returns -1 for a constant plane and the oracle agrees.
• The docstring now says the arctan2/compass conversion is duplicated from aspect._cpu on purpose, with the square tests guarding against drift.

The one remaining nit (the defensive has_dask_array() skip inside the cupy tests) is left as-is on purpose. It only matters on a CuPy-without-dask machine, and keeping it costs nothing while avoiding a confusing failure in that rare setup.

Oracle tests: 42 passed, 16 xfailed (numpy + dask; cupy paths skipped, no GPU here). flake8 clean on the added lines. No new findings.