Use CPU-only inference and enable hot reload

[CesarPetrescu]

Summary

• stub out flash_attn in its own module so CPU installs skip GPU extras
• drop accelerate to avoid pulling in GPU tooling
• force CPU model loading and drop CUDA handling
• install CPU-only torch wheels and enable uvicorn --reload
• mount source code for live reload in docker-compose
• symlink HuggingFace cache directories so rednote-hilab/dots.ocr loads correctly

Testing

• python -m py_compile invoice-extract/app/main.py invoice-extract/app/flash_attn_stub.py
• docker compose build --no-cache (fails: command not found)

---

https://chatgpt.com/codex/tasks/task_e_68a27c69d57883308e44b2149aef889c