fix(annotation): decouple partition_scope from dataset_id

[henrycgbaker]

Summary

• Introduces partition_scope: SafePathSegment on AnnotationSettings, defaulting to dataset_id (for backwards compatibility, but also as a sensible default)
• resolve_import_paths() and load_partition_manifest() now key off partition_scope; dataset_name() (Argilla naming) continues using dataset_id
• PartitionManifest.dataset_id renamed to partition_scope in Python, serialised as "dataset_id" via Pydantic alias — existing on-disk manifests load without migration
• import_records() API gains an optional partition_scope kwarg

Problem fixed

Two annotation configs sharing dataset_id: "" would share one partition manifest, so the first config to hit calibration_max_records would exhaust the cap for all subsequent imports in the same run — even across unrelated domains with different workspaces.

Backwards compat / future palns

• No changes required for existing callers. When partition_scope is not set, it defaults to dataset_id, preserving the exact previous behaviour. Existing partition.meta.json files on disk are unaffected (JSON key stays "dataset_id").
• Plan to refactor how datasets can be named with dataset_id (currently, has to be a suffix, but this is needless constraint now we have partition_scope)

Test plan

• 807 unit tests pass (uv run pytest tests/unit/)
• New integration regression test TestPartitionScopeDecoupling::test_independent_scopes_each_get_calibration — two configs with dataset_id: "", distinct partition_scope, calibration_max_records: 30; asserts scope B gets > 0 calibration records after scope A hits the cap
• Existing integration tests (tests/integration/test_annotation_calibration.py) unchanged in behaviour

[henrycgbaker]

Three things deliberately not fixed; capturing for future reference.

1. partition_scope: "" is indistinguishable from "not set" if not self.partition_scope treats an explicit empty string the same as an omitted field. A user who writes partition_scope: "" in YAML while having dataset_id: "DEM" will silently get partition_scope="DEM". Fixing this properly requires a None sentinel (SafePathSegment | None = None), which is a wider type change. Left as-is since an explicitly empty partition_scope with a non-empty dataset_id has no meaningful use case today.

2. Defaulting is one-way: partition_scope ← dataset_id, not vice versa:
   If a caller sets partition_scope="proj" without setting dataset_id, Argilla dataset names are unscoped (retrieval_production, no suffix). Intentional — the whole point of the decoupling — but a sharp edge. API docstring on import_records warns about this.

3. export_annotations() and teardown() don't expose partition_scope:
   Neither touches the partition manifest today, so no operational need. If either ever needs to correlate with manifest state, the kwarg will need to be added at that point.