Apply cell-size correction to planar aspect (#2760)

[brendancol]

Closes #2760.

What

Planar aspect() divided both Horn gradients by 8 without dividing by cellsize_x / cellsize_y, so it returned wrong directions for non-square cells. slope() already applied this correction; aspect did not.

• Thread cellsize_x / cellsize_y through the planar aspect path the same way slope() does: extract resolution in the public aspect(), pass it into the numpy/cupy/dask backend wrappers, and divide dz_dx by 8 * cellsize_x and dz_dy by 8 * cellsize_y in both the CPU (_cpu) and GPU (_gpu/_run_gpu) kernels.
• Square cells are unchanged because the factor cancels in the atan2 ratio, so the existing QGIS reference tests still pass.
• northness() and eastness() are fixed for free since they derive from aspect().

Backend coverage

The fix lives in the shared _cpu and _gpu kernels, so all four backends benefit: numpy, cupy, dask+numpy, dask+cupy.

Test plan

• New regression test: constant-gradient raster with xres=10, yres=1 now yields aspect 315.0 (was 275.7106), with matching northness (cos 315) and eastness (sin 315). Covers numpy and dask+numpy directly; cupy and dask+cupy variants run on GPU CI.
• Verified the new test fails on the unpatched kernel and passes after the fix.
• Existing aspect, slope, and geodesic-aspect suites pass (231 tests), confirming square-cell behavior is unchanged.

[brendancol]

PR Review: Apply cell-size correction to planar aspect (#2760)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• The aspect docstring still says the planar method "uses the classic Horn algorithm with uniform cell size". After this fix it handles non-uniform cells too. slope() carries the same stale wording, so leaving it keeps the two in sync; updating both would be a separate cleanup. Optional.

What looks good

• The fix mirrors the existing, tested slope() implementation: same 8cellsize_x / 8cellsize_y scaling, same get_dataarray_resolution call in the public function, same 1-element cupy arrays passed to the GPU kernel. Low risk because it copies a known-good pattern.
• The kernel orientation and compass conversion are untouched, so the QGIS reference tests still pass. Square cells are unchanged since the factor cancels in the atan2 ratio.
• All four backends are covered through the shared _cpu / _gpu kernels and the four wrapper functions; ArrayTypeFunctionMapping already dispatches to them.
• The regression test asserts the analytical Horn value (315.0) for a rectangular-cell raster rather than backend parity, and also checks northness/eastness. It was verified to fail on the unpatched kernel.

Notes checked, no action needed

• cellsize_y sign: calc_res derives yres from max-min of the y coords, so it is positive regardless of coordinate ordering, and the res attr is positive in practice. atan2 sign behavior matches slope's existing handling, so no new flip risk.
• Dask depth stays (1, 1) for the 3x3 kernel; boundary is forwarded unchanged. The cellsize values are bound via functools.partial, which is fine for map_overlap.
• Benchmarks already exist in benchmarks/benchmarks/aspect.py; no new function added, so no benchmark or README matrix change needed.

Checklist

• Algorithm matches the Horn cell-size correction used by slope()
• All implemented backends share the corrected kernels
• NaN handling unchanged (flat -> -1, edges -> NaN)
• Edge cases covered (existing degenerate-shape tests still pass)
• Dask chunk boundaries unchanged (depth 1, boundary forwarded)
• No premature materialization or extra copies
• Benchmark exists
• README feature matrix not applicable (no new function)
• Docstrings present (minor stale wording noted above)

[brendancol]

Follow-up review (after docstring commit)

The one nit from the first pass is fixed: the planar aspect docstring now says the gradients are scaled by the x and y cell sizes, so it no longer claims uniform cell size.

Nothing else changed in the code path. All 135 aspect tests still pass locally. No remaining findings.