Pre-resolve functional classification membership on the VA-Spec and NDJSON routes #657

The VA-Spec and NDJSON routes load every variant in each calibration range just to check which range one variant falls in. This issue replaces that with a single up-front query, so those routes stop loading thousands of rows they don't need.

Problem

_classification_contains_variant in lib/annotation/classification.py checks whether a variant falls in a functional classification. It prefers a pre-resolved set of classification ids, which is an O(1) check, and falls back to variant in functional_classification.variants. The fallback loads every variant in the range. Callers that don't pre-resolve pay that cost per classification, and routes eager-load ScoreCalibrationFunctionalClassification.variants to avoid an N+1 (for example routers/score_calibrations.py:749).

The cost lands on the VA-Spec and NDJSON routes: the study-result, functional-statement and pathogenicity-statement routes under /variants/{urn} and /mapped-variants/{urn}, and the score set pathogenicity NDJSON route.

Scope

• Pre-resolve containing classification ids for every caller, with one query against score_calibration_functional_classification_variants keyed by the variant ids in the request.
• Remove the ORM fallback from _classification_contains_variant, so the id set is required.
• Drop the selectinload(...variants) options that exist only for the membership check.

Acceptance criteria

• No caller of functional_classification_of_variant or pathogenicity_classification_of_variant loads ScoreCalibrationFunctionalClassification.variants.
• A query-count test on the score set pathogenicity NDJSON route asserts a constant number of queries as the variant count grows.
• Existing tests/lib/annotation coverage passes unchanged.