[ray] expose get_scheduler_actor_name so callers can look up the schedulers

[xyuzh]

Motivation

RayEngine names each SchedulerActor, but the name format is inlined in _create_scheduler_actor. A process outside the engine that wants handles to those schedulers — e.g. an RL trainer that supplied the placement group and wants to push weights straight into the scheduler actors — has no supported way to address them, and ends up scanning ray.util.list_named_actors() for substring matches.

Modifications

• Extract the name format into get_scheduler_actor_name(*, rank0_node_ip, dp_rank, pp_rank, tp_rank, port, bundle_idx) and call it from _create_scheduler_actor.
• Re-export it from sglang.srt.ray.
• Key the name by the engine's http port instead of _pg{hex}. The placement-group id is only knowable inside RayEngine, so an external caller cannot reproduce it; the port is assigned per host, so (node ip, port) uniquely identifies an engine and keeps engines co-located on one node distinct.

With this, a caller that knows the node ip, ranks, port and bundle indices it handed to RayEngine can rebuild the name and ray.get_actor() directly.

Note this does change the actor names RayEngine produces (_pg{hex} → _port{port}). Nothing in-tree reads the names — _create_scheduler_actor was the only place the format appeared.

Checklist

• Format the code
• No new dependencies

---

CI States

Latest PR Test (Base): ❌ Run #31641744701
Latest PR Test (Extra): ❌ Run #31641744995