fix: drop provider='hf-inference' (auto-route works for 72B)

fix: upgrade huggingface_hub + explicit hf-inference provider

fix: switch to Qwen2.5-72B-Instruct (7B not available on free inference API)

fix: LLM timeout + no-token fallback

Step 5: RAG with Qwen2.5

Initial deploy: FAQ bot with bge-small-zh