fix: lower token limits for CPU, streaming UI with multi-turn history, /chat instead of /ask

Initial commit: LFM2-700M FastAPI inference app