feat(ml-api): add reload-aware /predict response cache

[pavsoss]

Summary

/predict re-ran inference for every request, including identical repeats. This adds an in-process cache.

• A thread-safe TTL + LRU cache keyed by sha256(normalized_text) combined with the serving model version.
• /predict returns cached bodies on a hit and sets X-Cache: HIT|MISS; callers can bypass with Cache-Control: no-cache or ?fresh=1.
• A public GET /cache-stats reports hits/misses/size.

Why it's safe

Namespacing the key by the serving version means a model hot-swap invalidates the cache automatically — no stale outputs. Response bodies are unchanged; X-Cache is additive.

Test plan

Hit/miss behavior, TTL expiry, invalidation across a version bump, and the bypass header/param.

Closes #1008

[vercel]

@pavsoss is attempting to deploy a commit to the Aditya Sharma's projects Team on Vercel.

A member of the Team first needs to authorize it.

[Userunknown84]

Merge Conflict