Move LLM-backed operations to async job pattern

# Context

LLM-backed operations (label extraction, compliance evaluation) regularly take tens of seconds and can run a minute or more under `reasoning_effort=high`. Today they are issued as synchronous HTTP requests from the frontend, which has caused dropped connections to leave the UI stuck in a loading state with no recovery, and exposes the system to reverse-proxy timeouts. There is no way for a user to rejoin an in-flight request after a network blip.

# Outcome

LLM-backed operations run as async jobs whose result is retrievable independently of the originating HTTP connection. A dropped connection no longer loses or strands work.

# Acceptance criteria

- Label extraction and compliance evaluation are submitted as jobs that complete out-of-band from the request that started them.
- The frontend can observe a job's status (pending / running / succeeded / failed) and retrieve its result.
- A user whose connection drops mid-job can return to the label/compliance page and see the same job's current status and final result.
- Failed jobs surface a clear error in the UI; stuck/loading UI states from dropped connections are no longer possible.
- Existing UX (spinner during work, success/error toasts) is preserved or improved.