revert: go back to in-Space inference (free CPU on Space)

refactor: thin HF Inference API proxy for v5 models

chore: suppress transformers warnings via env var, clean up app.py

initial deployment