Phase B (live): Boltz-2 via Modal sidecar instead of ZeroGPU

Phase B: Boltz post-eval scaffold (graceful cpu-basic fallback)

Phase A: integrate LLM judge panel for hybrid scoring

fix: remove heavy boltz dep from requirements (lazy-imported at GPU runtime)

feat: add submission & scoring infrastructure (eval_scorer, dispatcher, boltz, queue, tasks) + fix gradio 5.x for Python 3.13

Upload requirements.txt with huggingface_hub