Add major world cities as bounding-box templates in from_template

[brendancol]

Closes #3533

Adds world cities to from_template, following the existing nyc pattern.

• 558 cities, generated from Natural Earth populated places and committed as static _CITIES data next to the country bboxes. No runtime network or geopandas dependency. Selection = every national capital + metros with POP_MAX >= 1.2M + a curated set of recognizable US secondary cities (Austin, New Orleans, Las Vegas, Portland, ...), since Natural Earth's POP_MAX/SCALERANK underrate US metros.
• Each city resolves to its UTM zone (EPSG:326xx north / 327xx south), picked from the centroid — a standard EPSG code, never a synthesized projection. Bounding box is a metro-scale lon/lat box (half-width scales with population) projected into that zone.
• Wired into _resolve between the curated regions and the country codes, so preserve='area'/'shape' work exactly as they do for nyc. No backend-dispatch changes.
• Name collisions: the larger-population city keeps the bare slug, the others take an _<iso2> suffix (e.g. hyderabad vs hyderabad_pk, birmingham UK vs birmingham_us).

from xrspatial import from_template
from_template("tokyo").attrs["crs"]        # 32654
from_template("austin").attrs["crs"]       # 32614
from_template("new_orleans").attrs["crs"]  # 32615

Backend coverage: numpy / cupy / dask+numpy / dask+cupy — cities reuse the same _make_data paths as every other template (verified locally on all four).

Test plan

• pytest xrspatial/tests/test_templates.py (56 passed)
• Integrity loop over all cities (keys, UTM-range CRS, ordered lonlat/bounds, ascii lowercase keys)
• Sample builds obey the array contract; pixel centers stay inside the bbox; southern-hemisphere build exercised
• UTM spot-checks (london 32630, tokyo 32654, southern -> 327xx), case-insensitivity, collision disambiguation
• Backend parity on dask+numpy / cupy / dask+cupy
• flake8 + isort clean on touched files
• Reconciled with main from_template: use CF Conventions metadata instead of crs_units/crs_name #3532 (CF grid-mapping metadata)

[brendancol]

Review: world-city bounding-box templates

Blockers

None.

Suggestions

• Discoverability. The unknown-name error points to "the templates reference" (templates.py:69-70), but neither templates.rst nor any public API lists the 519 city names — _CITIES is private. Country codes have the same gap, except ISO-3166 is a discoverable standard and an ad-hoc city list isn't. A follow-up that exposes a names listing (or a reference note on how to find them) would help. Not blocking.

Nits

• test_city_sample_builds (test_templates.py:187) builds only sorted(_CITIES)[:20], which is alphabetical and northern-heavy. A southern-hemisphere build is covered in practice because test_city_utm_spot_checks calls from_template("sao_paulo"), but adding one southern city to the sample loop would make that explicit.

What looks good

• Cities slot into _resolve between regions and countries (templates.py:49-55) and reuse _make_data unchanged, so all four backends work with no dispatch edits (checked locally on dask+numpy / cupy / dask+cupy).
• UTM-zone-from-centroid keeps every city on a standard EPSG code, and preserve='area'/'shape' fall through to the same nyc behavior for free.
• Data is committed static (no runtime geopandas or network), and the _CITIES header documents the source, the UTM rule, the buffer sizing, and the collision rule.
• test_city_registry_integrity checks the shape of all 519 entries, so a bad regeneration fails in CI instead of at call time.

Checklist

• CRS is a standard EPSG code (UTM), not synthesized
• All four backends consistent
• crs_units correct (metres for UTM cities)
• Integrity covered for every entry
• No dask/dispatch changes; reuses existing paths
• No premature materialization (static dict)
• No benchmark needed (allocation-only, matches existing from_template)
• README + reference + notebook updated
• Docstring updated with city example and collision rule

[brendancol]

Follow-up review

Changes since the first pass:

• Nit fixed: test_city_sample_builds now includes sao_paulo and sydney, so the southern-hemisphere (327xx) build path is exercised directly.
• Discoverability suggestion deferred to #3535 — a public listing API is a separate design decision (name, return shape, whether to include the 240 country codes).
• Merged origin/main and reconciled with #3532, which replaced crs_units/crs_name with CF grid-mapping attributes. The city test now asserts CF coord units (x.attrs["units"] == "m", standard_name == "projection_x_coordinate") and the user-guide cell prints grid_mapping_name instead of crs_name.

No new blockers. Local: 56 passed in test_templates.py, flake8 + isort clean, backend parity holds on dask+numpy / cupy / dask+cupy.