from_template: use CF Conventions metadata instead of crs_units/crs_name

[brendancol]

Closes #3531.

from_template was tagging its output with two non-standard attrs, crs_units and crs_name. This swaps them for CF (Climate and Forecast) Conventions metadata, the standard xarray and the geoscience stack already use.

Changes

• Drop crs_units. Units now live on the x/y coordinates where they belong. EPSG:4326 axes get degrees_east/degrees_north (was a non-CF 'degree'); projected axes get m, each with the matching CF standard_name (longitude/latitude, projection_x_coordinate/projection_y_coordinate).
• Replace crs_name with the CF grid-mapping keys grid_mapping_name (e.g. albers_conical_equal_area) and crs_wkt (full WKT, which carries the human-readable name). This lines up with the rest of the library, which keys off crs/crs_wkt.
• Source the CF keys from pyproj.CRS.to_cf(), kept best-effort: without pyproj the default path stays dependency-free and the keys are omitted, same as crs_name before. Equal Earth has no CF grid mapping, so grid_mapping_name is left off there and crs_wkt stands alone.

Backends

Metadata-only change; behaviour is identical across numpy / cupy / dask+numpy / dask+cupy.

Test plan

• test_cf_coordinate_units — CF units + standard_name on x/y, no crs_units
• test_cf_grid_mapping_attrs — grid_mapping_name + crs_wkt, no crs_name
• test_cf_grid_mapping_preserve_path — preserve path keys present
• test_grid_mapping_omitted_for_equal_earth — crs_wkt only for 8857
• test_cf_attrs_omitted_without_pyproj — graceful dependency-free fallback
• Full test_templates.py suite green (51 passed)

[brendancol]

PR Review: from_template — CF Conventions metadata

Blockers

None.

Suggestions

• Flat attrs vs. a real CF grid mapping (templates.py:325, :338-345). Strict CF puts grid_mapping_name on a separate scalar grid-mapping variable that the data variable points to via a grid_mapping attribute; here both grid_mapping_name and the coordinate standard_names sit directly on the data variable. That matches how this library already models CRS (flat attrs, crs/crs_wkt), so it's a reasonable call — but a strict CF reader like cf_xarray won't auto-detect it as a grid mapping. Worth a one-line docstring caveat, or a follow-up if fuller CF encoding is ever wanted. Not blocking.
• Narrower exception catch than before (templates.py:194-200). _cf_crs_attrs catches only ImportError; the old _crs_name also swallowed ValueError/TypeError. Every caller passes a curated/queried EPSG int, so from_epsg().to_cf() won't fail in practice — but if an unknown code ever reached here it would now raise instead of degrading to "no CF keys." Low risk given current call sites; flagging for awareness.

Nits

• Inconsistent wrapping (templates.py:339-345). The 4326 branch keeps .attrs.update(...) on one line (83 chars) while the projected branch wraps onto two. Both fit the 100-char limit; pick one style for symmetry.

What looks good

• Clean swap: no code outside the tests still reads crs_units/crs_name from a template, so nothing downstream breaks.
• crs_wkt lines up with the geotiff layer, which already documents crs/crs_wkt as the canonical, additive CRS keys.
• Test coverage is solid — both coordinate flavors, both grid-mapping branches, plus the two real edge cases (Equal Earth has no CF mapping, and the pyproj-absent fallback).

[brendancol]

Follow-up review (after fixes)

Re-reviewed the two follow-up commits. Disposition of the first-pass findings:

• Nit — inconsistent wrapping: fixed. Both branches of the coordinate-units block now use single-line .attrs.update(...) (all within the 100-char limit).
• Suggestion — flat attrs vs. CF grid-mapping variable: addressed by docstring. The Returns section now states the keys sit on attrs (the library's flat CRS convention), not a separate CF grid-mapping variable, so a strict CF reader won't auto-detect them.
• Suggestion — narrower exception catch in _cf_crs_attrs: dismissed. Every call site passes a curated or pyproj-queried EPSG int, so from_epsg().to_cf() can't fail in practice. CLAUDE.md asks for no error handling on unreachable paths, so I left the catch scoped to ImportError (the only real branch — the dependency-free fallback).

flake8 clean; test_templates.py green (51 passed). No blockers outstanding.