Recompute res attribute for downsampled preview() output

[brendancol]

Closes #3569.

preview() downsamples a raster and rebuilds the x/y coordinates for the smaller grid, but it copied the input's attrs verbatim. So attrs['res'] kept describing the original full-resolution grid, and any downstream code reading res for pixel-size math or georeferencing got the wrong value.

Change

• Add _res_from_coords() in xrspatial/preview.py to read the pixel size off the output coordinate spacing (positive magnitude, matching how templates and terrain store res).
• Recompute res once, right before preview() returns, after the second-pass refinement has settled the final coordinates. This covers every method (mean, median, max, min, nearest, bilinear) and all four backends in one place.
• Only touch res when the input already had one, and only when both axes have at least two coordinates. A raster without res stays without it; a no-op passthrough keeps its original.

Backends

numpy, cupy, dask+numpy, dask+cupy — verified on all four. The one combination not covered is dask+cupy + bilinear, which already fails before this change in _zoom_block (scipy zoom on a cupy block); that is a separate pre-existing bug, untouched here.

Test plan

• res matches the returned array's coordinate spacing for all six methods
• source-grid res does not survive a downsample
• input attrs is not mutated
• passthrough (no downsample) keeps the original res
• input without res stays without res
• dask path through snap + refinement
• cupy path
• full test_preview.py suite: 59 passed

[brendancol]

PR Review: Recompute res attribute for downsampled preview() output

Blockers

None.

Suggestions

• preview() docstring (xrspatial/preview.py:350-353) still says the output has "updated coordinates" but not updated resolution. Worth a half-sentence so readers know res now tracks the downsampled grid.

Nits

• _res_from_coords (xrspatial/preview.py:285-300) reads spacing from the first two coordinates only. That is exact for the uniform grids preview produces (interp over a linspace, or a strided isel), so it is fine here — just noting it assumes uniform spacing.

Observations (not action items)

• nearest reports a slightly different res than mean/bilinear (5.0 vs 5.115 in my checks). That traces to how each path builds coords: nearest strides (spacing = factor x original) while the reduce/bilinear paths interpolate across the full original extent. The PR faithfully reflects whatever coords each path emits; the coord-construction difference predates this change.
• If an input carries res but no x/y coords, the recompute is skipped and res stays stale. That input is self-contradictory (res describes coord spacing) and there is no coordinate to measure, so leaving it alone is the right call.

What looks good

• One recompute point after the second-pass refinement covers all six methods and all four backends, instead of patching each helper.
• Builds a fresh attrs dict, so the input is not mutated (covered by a test).
• Only touches res when it was present and measurable; absent-res inputs and no-op passthroughs keep their original attrs, both tested.
• Tests assert against the returned array's real coordinate spacing rather than a hardcoded number, and confirm the stale source value does not survive.

Checklist

• Result reflects the actual output coordinates
• All implemented backends produce consistent results (numpy/cupy/dask+numpy/dask+cupy; bilinear+dask+cupy excluded by a pre-existing unrelated bug)
• No input mutation
• Edge cases covered (passthrough, absent res, dask refinement)
• No premature materialization or unnecessary copies
• Docstrings present (minor wording suggestion above)
• No README/benchmark change needed (bug fix, no API or perf change)

[brendancol]

Follow-up review (after docstring commit 1227f85)

Re-checked the diff after the follow-up commit.

• Suggestion from the first pass (document the res recompute) — fixed. The preview() Returns section now states attrs['res'] is recomputed from the output spacing when the input carried a res.
• Nit (first-two-coords spacing) and the two observations — left as written; they describe correct, intended behavior, not defects.

No new findings. flake8 clean on preview.py; the 17 res tests still pass. No blockers.