Add global-projection templates to from_template (web_mercator, wgs84/latlon, equal_earth)

[brendancol]

Closes #3549

Adds whole-world templates to from_template, each resolving to a real EPSG code:

• web_mercator -> EPSG:3857 (WGS 84 / Pseudo-Mercator)
• equal_earth -> EPSG:8857 (Equal Earth, the equal-area option)
• wgs84 and latlon -> aliases for the existing EPSG:4326 world grid (single source of truth via a small _REGION_ALIASES map)

Robinson, Gall-Peters, and Mollweide were requested but only have ESRI codes, not EPSG. Storing those as a bare int would be ambiguous with EPSG and would break the CF metadata path, so they are left out; equal_earth covers the equal-area case with a real EPSG code. The reasoning is written up in the issue.

The new names go through the same backend dispatch as the rest of from_template, so numpy, cupy, dask+numpy, and dask+cupy all work with no new backend-specific code. The dask path stays lazy (da.full).

Test plan

• pytest xrspatial/tests/test_templates.py (80 passed, including the cupy and dask backends that run on this box)
• Contract checks for all four names (dims, all-NaN, float32, crs int, north-up coords)
• wgs84/latlon produce the same grid and attrs as world
• Case-insensitive lookup, exact-resolution honoring, metre coords within bounds
• CF grid-mapping: 3857 and 8857 have no grid_mapping_name, so crs_wkt carries the name
• Verified the web_mercator dask grid stays lazy

[brendancol]

PR Review: Add global-projection templates to from_template

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• xrspatial/_template_data.py: the web_mercator and equal_earth entries carry area_epsg=8857 / shape_epsg=3395, so from_template('web_mercator', preserve='shape') will reproject a global template into yet another projection. That combination works but is odd, and nothing tests it. Either add a one-line build test or drop the keys if preserve-on-global is out of scope.

Nits (optional improvements)

• xrspatial/_template_data.py: the Web Mercator latitude limit is written as 85.051129, but the exact value is 85.0511287798. It only matters on the rarely-used preserve path, so it is cosmetic. The nearby comment cites the precise definition, which makes the rounded literal look slightly off.

What looks good

• wgs84/latlon reuse the world grid through _REGION_ALIASES rather than copying the bounds, so there is one source of truth.
• All four projections resolve to real EPSG ints, matching the project convention. The ESRI-only projections (Robinson, Gall-Peters, Mollweide) are documented as deliberately left out.
• No new backend code: coverage comes from the existing _make_data dispatch, and the dask path stays lazy via da.full.
• Tests cover the array contract, the alias-equals-world check, case-insensitive lookup, exact resolution, bounds containment, and the CF grid-mapping behavior for 3857/8857.

Checklist

• Algorithm matches reference (EPSG codes and projected bounds verified with pyproj)
• All implemented backends produce consistent results (shared dispatch)
• NaN handling is correct (NaN-filled, asserted)
• Edge cases covered (case-insensitive, exact resolution, bounds)
• Dask chunk boundaries handled (lazy da.full, no overlap needed)
• No premature materialization
• Benchmark not needed (no compute kernel)
• README feature matrix updated
• Docstrings present and accurate

[brendancol]

Follow-up review (after commit b1719cb)

Both items from the first pass are resolved.

• Suggestion (preserve on global templates): kept the area_epsg/shape_epsg hints, since they match the world entry and steer preserve='shape' to World Mercator (EPSG:3395) instead of a stray UTM zone for the (0, 0) centroid. Added test_global_preserve_picks_world_projection to lock that in for both web_mercator and equal_earth.
• Nit (latitude limit): the Web Mercator lonlat band now uses the exact 85.0511287798, matching the comment.

No new findings. Tests: 82 passed locally (numpy, cupy, dask backends).