Consolidate the EOPF Zarr product schema through pydantic on Zarr V3 #258

Problem

Every input product model is built on the Zarr V2 pydantic model, while CPM 3.0.0 now emits Zarr V3. So validation silently stops working on current products, and the code works around it by hand.

data_api/s2.py        -> from eopf_geozarr.pyz.v2 import ArraySpec, GroupSpec
data_api/s1.py        -> from eopf_geozarr.pyz.v2 import ArraySpec, GroupSpec
data_api/s3_olci.py   -> from eopf_geozarr.pyz.v2 import ArraySpec, GroupSpec

pyz/v3.py already exists and provides GroupSpec / ArraySpec over pydantic_zarr.v3. Nothing in src/ imports it. The scaffolding is there; the models were never ported.

Symptoms accumulating in the meantime

• Cpm geozarr eodc changes #247 adds _is_sentinel2_dataset_v3() — a hand-rolled structural check (set(group.keys()) == {"measurements","quality","conditions"}, then look for measurements/reflectance). Its own docstring says why: "Sentinel2Root ... is defined against the Zarr V2 pydantic model only; there is no V3 counterpart yet." Reasonable as a stopgap, but it is a schema expressed as imperative code.
• CRS extraction from other_metadata (s2_converter.py:41-60) reads horizontal_CRS_code / horizontal_crs_code, coerces "EPSG:32632" / "32632" / 32632, and swallows failures in a broad except Exception. Two spellings and three representations of one field, parsed by hand — exactly what a model should own.
• Earlier in the CPM writer work, is_sentinel2_dataset() and is_sentinel3_olci_dataset() both had to be defensively guarded because they validate against V2 models and fail on V3 stores.

Each workaround is individually defensible. Together they mean the product schema now lives in several places, none of them the model.

Proposal

Consolidate the schema of the EOPF Zarr product through pydantic, on pyz.v3:

1. Port Sentinel2Root, Sentinel1Root and Sentinel3OlciRoot to pyz.v3, keeping the V2 models only as long as V2 inputs must still be read.
2. Model the CPM root metadata we actually depend on — at minimum the other_metadata CRS field, with its spelling and type variants handled once, in the model.
3. Replace the hand-rolled structural detectors (_is_sentinel2_dataset_v3, and the guards around the existing is_*_dataset helpers) with model validation, so "is this an S2 product?" has one answer.
4. Decide what the models own versus what stays imperative, and write that down — the point is one source of truth for the product schema, not maximal coverage.

Notes

• Related, but separate: Specify which CPM root metadata is carried through to GeoZarr output (stac_discovery is stale by construction) #257 covers which CPM root metadata we carry through to the output. This issue is about typing what we read.
• @d-v-b — flagging you as the author of the pydantic model layer; happy to discuss scope before anyone starts.