[Shine-neko]

Second step of the restart policy design (documentation/design/restart-policy.md).

• Docker and Podman workers are reconciled by the scheduler through a new instance driver interface (observe / start / stop / discard) and the restart policy: they back off with full jitter (10s up to 5m) and are never abandoned unless on_exhaustion = "fail". The counter resets after stable_after (10m).
• Restart state (next attempt, running since, last termination with a tail of its output) is persisted, so a server restart no longer resets backoffs.
• New [server.restart] section, validated at startup.
• crash_loop_back_off and insufficient_resources are reconciled again; liveness kills count as failures.
• Jobs and the containerd, Cloud Hypervisor and Firecracker runtimes keep their apply path and the 5-attempt budget for now, with retries spaced on the persisted backoff. RetryBackoff is removed.
• Docker e2e suite and documentation updated.