Wire up InferenceController and RolloutExecutor and make the controller API async

[fzyzcjy]

Part of #1837

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[guapisolo]

Could we pull two small, dependency-light fixes forward and place them immediately after #1843, so this prefix can merge without waiting for the full refactor stack?

1. Extract the minimal process-seam fix from Cut the eval fleet along the process seam #2614: keep the controller-side and executor-side EvalFleet state separate, and pass only data such as EvalFleetInfo / EvalFleetPin across the process boundary. This closes the serialized-copy/state-divergence bug introduced here. This should be described as a process-boundary correctness fix only: Delete the rollout health monitor ahead of its rewrite #1847 and Temporarily remove vanilla fault tolerance for the eval fleet #2497 intentionally remove the old eval-fleet FT path, so it must not claim that dedicated eval recovery is restored.
2. Preserve replay compatibility for persisted audit JSONL with component="rollout_manager", either through a legacy alias or an explicit migration to rollout_executor. Let every miles process name itself as the component it is #2503 still accepts only rollout_executor, so older records remain unreadable or get dropped.

Ideally these are minimal follow-up commits/PRs directly on top of #1843, and we merge that smaller safe prefix before the remaining #1837 stack.