perf: reduce heatmap from 256 to 16 model inferences + fix gunicorn timeout/workers

deploy: clean initial commit for Hugging Face Spaces (no binary blobs)