[Serve][4/4] Add build_openai_applications for multi-model direct streaming

[celinaky]

Why this is needed

This PR adds a public (alpha) builder that supports multi-model direct streaming via a multiple model applications supervised by a router application and a control application.

Summary

ray.serve.llm.build_openai_applications(llm_serving_args, *, name="llm", route_prefix="/") -> List[RunTarget]

For N configured models, the builder returns N + 2 application targets: one per model application, a router application, and a control application.
Each model application uses the existing single-model direct-streaming topology.

Example usage

from ray import serve
from ray.serve.llm import LLMConfig, build_openai_applications

apps = build_openai_applications({"llm_configs": [qwen_config, llama_config]})
serve.run_many(apps)

POST /v1/chat/completions {"model": "qwen"}   → router → streamed from a qwen LLMServer replica
GET  /v1/models                               → router → control application
POST /models/qwen/v1/chat/completions         → qwen application directly (skips the router)

Test plan

• pytest .../routers/test_builder_openai_applications.py: builder unit tests
• RAY_SERVE_ENABLE_HA_PROECT_STREAMING=1 pytest.../test_openai_applications_direct_streaming.py
• pre-commit run on changed files