rasterize/kde/line_density: set output attrs['res'] when grid resolution changes (#3571)

[brendancol]

Closes #3571

Tools that build an output grid at a different resolution than their input now write the correct res into the output attrs. get_dataarray_resolution() reads attrs['res'] before deriving cell size from coordinates, so a missing or stale res was handing the wrong cellsize to downstream distance/area ops like slope, aspect, and proximity.

• rasterize(): the reshape path (caller overrides bounds/width/height/resolution, or rasterizes with no like at all) now records the freshly-built grid's cell size as attrs['res'] instead of leaving it absent. The reuse-like-coords path still keeps the template's res/transform unchanged. A stale inherited transform is still dropped on reshape.
• kde() / line_density(): when template is omitted, the output now carries attrs['res'] = (abs(dx), abs(dy)) for the grid built from x_range/y_range.

resample() already did this on every path; this aligns the remaining resolution-changing tools with it. preview() has the same class of issue and is handled in a separate PR.

Backend coverage: the rasterize change covers numpy, cupy, dask+numpy, and dask+cupy (attrs are built backend-independently, and the #2251 test class exercises all four). The kde/line_density change is on the eager template=None construction path.

Test plan:

• Updated the issue rasterize: stale transform/res attrs propagated from like when grid is overridden #2251 rasterize tests (and the no-like rectangular-pixel test) to assert the refreshed res value instead of asserting res is absent.
• Added TestKDEResolutionAttr covering kde/line_density res from range, res-matches-coords, and template-supplied res.
• test_kde.py, test_rasterize.py, and test_rasterize_coverage_2026_05_21.py pass (340 passed, 2 skipped).

[brendancol]

PR Review: rasterize/kde/line_density: set output attrs['res'] when grid resolution changes (#3571)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• xrspatial/kde.py:676, 805 — For a single-column or single-row output (width=1 or height=1), dx/dy comes out as span / max(cols-1, 1) = span, so res reports the full extent on that axis (e.g. width=1 over x_range=(0,10) gives res_x=10 against a single x coord at 0). That's a degenerate grid and the value just echoes the existing dx/dy binning, so it isn't wrong, just meaningless. Not worth special-casing unless single-pixel KDE grids are a real use case.
• test_kde.py — test_kde_res_matches_coords checks res-vs-coords and get_dataarray_resolution for kde; line_density only gets the from-range assertion. A matching res-vs-coords check for line_density would be symmetric, even though the construction path is the same.

What looks good

• The reuse-like-coords path is untouched, so the bit-identical-template contract and the existing res/transform preservation still hold. Only the genuinely-reshaped path changes.
• abs() on px/py/dx/dy keeps res positive when a caller passes a descending x_range/y_range (checked: x_range=(10,0) yields res (1.0, 1.0), not negative).
• res matches the coordinate-derived cellsize under both conventions: rasterize uses cell-center coords, kde uses an endpoint-inclusive linspace, and res equals the coord spacing in each.
• Stale transform is still dropped on reshape, so res and transform can't disagree.
• The #2251 test class was updated rather than deleted, so the regression coverage survives and now pins the stronger contract: res is present and correct instead of just absent.

Checklist

• res order is (x_res, y_res), consistent with resample() and get_dataarray_resolution()
• All four backends covered for rasterize (attrs built post-dispatch, backend-independent; #2251 dask/cupy tests assert res)
• NaN handling unaffected (metadata-only change)
• Edge cases covered (rectangular pixels, no-like, like-reshape, template-supplied)
• Dask chunk boundaries unaffected (no compute path change)
• No premature materialization or copies introduced
• No benchmark needed (metadata-only)
• README feature matrix not applicable (no new function, no backend change)
• Docstrings updated for kde, line_density, and rasterize Returns

[brendancol]

Follow-up review (#3571)

Re-reviewed after the follow-up commit (939d358a). The commit only adds a test, no production change, so the prior review's findings still stand. Disposition of the two nits:

• Nit 2 (line_density res-vs-coords test) — fixed in 939d358a via test_line_density_res_matches_coords, mirroring the kde check.
• Nit 1 (single-cell res reports the full extent) — dismissed. For a 1-cell axis, res mirrors the dx/dy the kernel already uses for binning, so it stays consistent with how the grid was built. There's no sensible alternative value, and special-casing a degenerate config would add code for no real gain.

No blockers or suggestions. Good to merge once CI is green.