rasterize: honor resolution= exactly when bounds don't divide evenly

[brendancol]

Closes #2573.

Summary

• rasterize(..., resolution=R) used to compute width/height via ceil(extent / R) but rebuilt output coords over the original bounds, so the actual cell size shrank to fit. bounds=(0,0,1,1), resolution=0.3 returned a 4x4 raster with cell size 0.25 instead of 0.3.
• After resolving width/height, final_bounds is rebuilt so the grid anchors at (xmin, ymax) and extends right / down by exactly width * x_res / height * y_res. Matches rasterio.transform.from_origin(west, north, x_res, y_res).
• When the original bounds already divide evenly, the new values equal the originals to float precision, so existing callers see no change.

Backend coverage

Fix lives upstream of backend dispatch, so all four backends pick it up: numpy, cupy, dask+numpy, dask+cupy. Cross-backend parity is pinned by the new test class.

Test plan

• New regression test for the headline 4x4 / 0.25 example
• Divides-evenly case still passes (no behavior change)
• Asymmetric (x_res, y_res) tuple with non-divisible bounds
• Polygon burn within original bounds is preserved when the grid expands
• Cross-backend parity (numpy / cupy / dask+numpy / dask+cupy)
• Full existing rasterize test suite passes (505 tests)

[brendancol]

PR Review: rasterize resolution= fix (PR #2597)

Blockers

None.

Suggestions

• xrspatial/tests/test_rasterize_resolution_exact_2573.py: no test combines like= with resolution= against a template whose bounds don't divide evenly. The existing test_resolution_override_strips_stale_grid_attrs at test_rasterize.py:2096 uses resolution=0.5 on a 10x10 template, so bounds divide evenly and the new overshoot path isn't exercised. A test with like= (10x10 template) plus resolution=0.3, asserting the output cell size is 0.3, would close that gap.
• xrspatial/rasterize.py:3239-3241: the inferred-bounds + resolution path also picks up the fix. When a GeoDataFrame's total_bounds supplies non-divisible bounds and the caller passes resolution=, the grid now extends past the geometry envelope to honor the cell size. No test pins this. A small case with an irregular geometry, no explicit bounds=, and resolution=0.3 would lock in the inferred-bounds branch alongside the explicit-bounds branch.

Nits

• xrspatial/rasterize.py:3239-3241: worth pinning the get_dataarray_resolution(result) side-effect. After the fix, that helper reports the cell size the caller actually asked for. A one-line assert against it in the existing test documents the downstream win for slope / aspect / proximity, which previously inherited the shrunk cellsize.

What looks good

• Only the resolution= branch is touched. Explicit width/height and like=-only paths keep their behavior byte-for-byte.
• Anchoring at (xmin, ymax) matches rasterio.transform.from_origin, which is what GDAL / rasterio users will expect.
• width * x_res == xmax - xmin to float precision when bounds divide evenly is why the existing tests still pass without modification. The inline comment and CHANGELOG capture this.
• Cross-backend parity is pinned for numpy, cupy, dask+numpy, and dask+cupy, since the fix lives upstream of dispatch.
• Docstring update tells users what to expect when bounds don't divide evenly and points at the matching rasterio API. Better than silently changing behavior.
• 505 existing rasterize tests pass unchanged.

Checklist

• Algorithm matches the rasterio.transform.from_origin convention
• All four backends produce consistent results
• NaN / fill semantics unchanged
• Edge cases covered: divides-evenly, doesn't-divide-evenly, asymmetric tuple, geom inside original bounds with expanded grid
• No dask chunk boundary concerns; the change is in bounds resolution, not in the rasterize kernels
• No new materialization or copies
• [n/a] No new public function, so no benchmark or README feature matrix entry needed
• Docstring updated for the new contract

[brendancol]

PR Review (follow-up): rasterize resolution= fix (PR #2597)

Re-reviewed after c7eeacad. All three prior findings are now addressed.

Disposition of prior findings

• Suggestion 1 (like= + non-divisible resolution=): fixed by TestLikePlusResolutionOvershoot (10x10 template + resolution=0.3 -> 34x34 grid with exact cell size).
• Suggestion 2 (inferred-bounds + resolution=): fixed by TestInferredBoundsPlusResolution (no explicit bounds=, geometry envelope drives final_bounds).
• Nit (get_dataarray_resolution side-effect): fixed by test_get_dataarray_resolution_matches_request.

Blockers

None.

Suggestions

None.

Nits

None.

12/12 regression tests pass. Existing 505-test rasterize suite still green. No further findings.