Honor requested resolution exactly in from_template() (#3494)

[brendancol]

Summary

from_template(name, resolution=r) now honors the requested resolution exactly. The previous code held the template's bbox fixed and recomputed the realized cell size from it, so res drifted when the extent was not a whole number of cells (e.g. from_template('nyc', resolution=10) returned res=(10.0, 10.000757)).

• Anchor the lower-left corner and nudge the far edges (right, top) to an exact multiple of the cell size, so attrs['res'] matches the requested resolution.
• This matches the reproject grid path (_compute_output_grid in reproject/_grid.py).
• Pixel centers still fall inside the original registry bbox; the far edges move by less than half a cell.

Closes #3494

Backends

Backend-agnostic: the change is in the shared grid math before the array is allocated, so numpy, cupy, dask+numpy, and dask+cupy all get the same exact res.

Test plan

• from_template('nyc', resolution=10).attrs['res'] == (10.0, 10.0)
• tuple resolution honored exactly
• coord spacing equals the requested resolution
• pixel centers stay within the registry bbox
• preserve path honors resolution exactly
• full test_templates.py suite passes (44 tests)

[brendancol]

PR Review: Honor requested resolution exactly in from_template() (#3494)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• The new exact-resolution tests only cover projected-CRS regions (nyc, conus) and the preserve path. Country codes come back in EPSG:4326 and run through the same nudge math, but nothing pins exact res for a 4326 template. A one-liner like from_template('FRA', resolution=0.25).attrs['res'] == (0.25, 0.25) would close that gap. (xrspatial/tests/test_templates.py)

Nits (optional improvements)

• The dask backend's exact res is covered indirectly by test_dask_numpy_backend (it asserts attrs == numpy ref). An explicit assertion would document the intent, but it's optional. (xrspatial/tests/test_templates.py:139)

What looks good

• The fix mirrors the existing reproject grid path (_compute_output_grid in reproject/_grid.py lines 457-459), so the two paths now agree on how resolution is honored.
• The change sits in the shared grid math before _make_data, so all four backends get identical coords and res with no per-backend code.
• Pixel centers stay inside the original registry bbox: the center sits res/2 inward, which absorbs the half-cell nudge. Checked both axes.
• I checked the EPSG:4326 country path for latitude/longitude overflow from the nudge. Latitude never gets near 90 (the northernmost bbox top is ~83.6 for GRL), and the lon>180 values come from pre-existing antimeridian-wrapped bboxes, not from the nudge. No new invalid-coordinate risk.
• Docstring updated to describe the nudge and the lower-left anchor; autosummary picks it up.

Checklist

• Algorithm matches the reproject grid convention
• All four backends produce consistent results (shared pre-allocation math)
• NaN handling unchanged (fill path untouched)
• Edge cases covered (tuple res, centers within bbox)
• No dask chunk concern (change is before array allocation)
• No premature materialization
• No benchmark needed (constant-time arithmetic)
• No README change needed (no new function, no backend change)
• Docstring present and accurate

[brendancol]

Follow-up review (after 0dc8a4e)

Both findings from the first pass are addressed:

• Suggestion (4326 exact-res test): fixed. Added test_country_resolution_honored_exactly, asserting from_template('FRA', resolution=0.25).attrs['res'] == (0.25, 0.25).
• Nit (explicit dask exact-res): fixed. test_dask_numpy_backend now asserts the dask+numpy path returns res == (10.0, 10.0).

Full test_templates.py suite passes (45 tests). No new findings.