Build bulk response rows without validating each through Pydantic #915

Summary

Whole-set endpoints build each row as a dict and then validate it through a Pydantic view model before serializing, for example VariantDetail.model_validate(detail) in the /variant-details stream. On a 500k-variant score set that validation runs half a million times on data the server just built. This issue serializes the built rows directly and moves the shape guarantee into tests.

Scope

• Serialize rows straight to JSON in the /variant-details stream (_stream_score_set_variant_details, routers/score_sets.py), the lean view (get_lean_score_set_variants), the /annotated-variants/* streams, and wherever else the profiling shows validation cost.
• Keep the Pydantic models as the OpenAPI schema and as the single-object routes' response models.
• Add a test per endpoint that validates every emitted row against its model on a seeded score set.

Acceptance criteria

• The listed endpoints don't call model_validate per row.
• Each endpoint's output validates against its view model in tests.
• Each endpoint's output is byte-identical to the validated output on the test score set.
• Time per 10k rows drops by at least the validation share measured in the profiling issue.