Switch to vLLM with custom stopping criteria

[lee101]

Summary

• add vLLM based inference helper with min-probability and sentence stopping
• integrate vLLM into the FastAPI server when available
• make tests/conftest resilient if httpx is missing
• add basic unit test for vLLM inference

Testing

• pytest -q (fails: ModuleNotFoundError: No module named 'cachetools')

---

https://chatgpt.com/codex/tasks/task_e_683fe5b06bf083338fb1ba3540b415dc