Add LLM reachability check to backend prestart

# Context

The current backend prestart sequence (`backend/scripts/prestart.sh`) validates DB and MinIO storage before the `backend` service starts, but not Azure OpenAI. Misconfigured LLM credentials, wrong endpoint, wrong API version, or a removed deployment currently slip through — the app boots clean and only fails when the first label extraction or compliance evaluation request lands.

# Outcome

Backend prestart fails fast on Azure OpenAI misconfiguration, consistent with how DB and storage are validated today.

# Acceptance criteria

- Prestart sequence verifies Azure OpenAI reachability before the `backend` service starts.
- The probe does not consume tokens (e.g. list deployments / list models, not a completion call).
- Failure surfaces a clear error identifying Azure OpenAI as the failing dependency.
- Failure exits non-zero so the `prestart` compose service does not complete successfully and the `backend` service does not start.