Launch the sgl and miles routers through RayWorkerManager

[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]

[P3] Keep debug_train_only free of rollout services

Codex review: launch_worker_manager() runs before the later InferenceController.init() guard, so compute_specs() still emits router workers in --debug-train-only mode. This does not affect training correctness, but it unnecessarily starts a router actor and binds ports despite the documented train-only contract; could we return no router specs when args.debug_train_only is set?

[guapisolo]

[P2] Preserve the MILES_HOST_IP router override

Codex review: this now derives the router host only from CommandActor._get_node_ip(), whereas the previous path used get_host_info() and honored MILES_HOST_IP. This is non-blocking when Ray’s node IP is already usable, but deployments that set the override for a container/pod address will bind and publish the Ray-detected IP instead. The behavior remains present through #2176 and #2649; could the router spec carry an explicit bind/advertised-host override?

[guapisolo]

Codex stack review: #2054 completes the router/session placement transition, while #2053 and #2057 close the source-test gaps; all land before safe-merge checkpoint #2176 and remain intact through checked tip #2649. The two inline comments are non-blocking follow-ups.