Clean up GeoZarr output metadata: proj/spatial only, well-formed attributes #350

Context

The independent-reader review in #349 shows duplicate and malformed metadata in the S2 L2A output of our GeoZarr engine. The Zarr Sample Service uses this engine in CPM, so these problems are in the published products.

Decisions

1. proj, spatial and multiscales are the only normative georeferencing.
2. The output makes no CF claim. We do not write a root Conventions attribute. CF is out of scope of the minispec (docs/geozarr-minispec.md). CF alignment is a spec topic (CF Convention Alignment zarr-developers/geozarr-spec#120).
3. The GeoZarr V0 grid_mapping and spatial_ref are obsolete.
4. CF-style attributes that we keep (flag_*, valid_*, scale_factor, add_offset) must be well-formed.
5. _FillValue does not change. xarray needs it for Zarr v3 (fix: ensure fill_value for zarr v3 arrays round-trips pydata/xarray#11345).

Tasks

Downstream check

We searched titiler-eopf, titiler-openeo, data-pipeline, eodash-assets and the data-model tests. No code reads flag_meanings, short_name, the dtype attribute or Conventions. titiler-eopf copies proj:* and spatial:* from the group to the array (titiler/eopf/reader.py), so it does not need grid_mapping. We did not check eopf-stac.