serving: release_id keys audits, quarantine and the READY pool; gateway routes by model

[anderdc]

One model can have several blessed runtimes, and phase 1 assigns gpu -> release. Keying the audit window, quarantine, READY pool and served requests by model_id collides the moment two releases serve the same model — so everything the validator tracks moves to release_id. It defaults to model_id, so today's single-release loadout behaves identically.

The bar this is aimed at: adding a second model (a 27B dense on the same sparkinfer runtime) should be a loadout entry plus a reference, not a refactor.

• loadout — release_id on ServingRelease, unique across the loadout; get() resolves a release_id then a model_id. An optional audit block per release overrides the three logprob bands (a non-deterministic runtime ships its own; the constants stay the default).
• gateway — /v1/models lists every release; model in the request body picks one (404 when it is not served); the request is audited against that release's reference.
• miner — InferenceSynapse carries release_id, and a miner serving another release refuses the call rather than answering for a release it was not blessed for.
• window — strikes counted per (hotkey, release) and persisted, so escalating quarantine becomes a policy constant rather than a schema change.
• store — audit_values / quarantine rename model_id -> release_id in place and gain strikes; existing rows carry over under their old key.

Pairs with entrius/gittensor-db#NN (release_id columns) and entrius/das-gittensor#NN — the DB columns should land first.

CI: ruff, format, pyright, vulture, 742 tests green.