chore(docker): pin flash-attn prebuilt wheel to cp311 torch2.7 cu126; remove pip fallback

chore(docker): try pip install flash-attn fallback when wheel tag mismatches

chore(docker): create /data dirs and chown for model/cache; set OMP_NUM_THREADS=1

feat(api): fast FastAPI app + model loader refactor; add mock mode for tests\n\n- Add pyproject + setuptools config and console entrypoint\n- Implement enhanced field extraction + MRZ heuristics\n- Add response builder with compatibility for legacy MRZ fields\n- New preprocessing pipeline for PDFs/images\n- HF Spaces GPU: cache ENV, optional flash-attn, configurable base image\n- Add Make targets for Spaces GPU and local CPU\n- Add httpx for TestClient; tests pass in mock mode\n- Remove embedded model files and legacy app/modules

Docker: guard flash-attn install by Python ABI tag; skip on mismatch

Enable CUDA 12.6 runtime, default flash-attn on HF, add huggingface_hub/PyMuPDF, make flash-attn configurable

Debian Trixie compat: use libgl1, libxrender1; drop libatlas-base-dev

Add OpenGL and OpenCV dependencies for headless operation

Fix Docker build: replace deprecated libgl1-mesa-glx with libgl1-mesa-dri

Prepare for Hugging Face Spaces deployment

Initial commit