Downloading Qwen2.5-1.5B + MiniLM before app.launch() blocks the
Gradio server from starting within HF's ~5 min startup timeout.
Models are already lazily loaded on first request via model_loader.py.
Remove Docker files; fix CPU-only torch for Gradio SDK space
Drop Dockerfile, docker-compose.yml, .dockerignore.
Add --extra-index-url for PyTorch CPU wheel so HF installs the
lightweight CPU build instead of the 2.5 GB CUDA package.
