Add debug logging for received image data

Fix base64 data URI detection: use startsWith instead of comma check

Change custom API path to /custom_api/ to avoid Gradio /api/ conflict

Use FastAPI + mount_gradio_app for custom /api/digitize_base64 endpoint

Add base64 API endpoint for programmatic access

Fix RoIAlign/RoIPool output_size to always be a tuple

Implement full CPU forward for MultiScaleDeformableAttention

Import MultiScaleDeformableAttention in ops __init__ for registry

Register MultiScaleDeformableAttention in mmcv ATTENTION registry

Add sigmoid_focal_loss, deform_conv2d, get_onnxruntime_op_path to shim

Add missing mmcv.ops stubs: RoIPool, DeformConv2d, etc.

Fix apt package: libgl1-mesa-glx -> libgl1 for Debian Trixie

Fix PyTorch install URL: use --extra-index-url for CPU wheels

Fix Dockerfile: extract mmcv.ops shim to separate Python file

Fix: inline MMCV_WITH_OPS=0 in RUN + --no-build-isolation

Use mmcv-full with MMCV_WITH_OPS=0 (skip C++ extension build)

Switch to Docker SDK for reliable mmcv-full build

Fix requirements: use --extra-index-url for torch, remove gradio dup

Fix build: pin torch 1.13 + mmcv-full pre-built wheel

Initial deploy: AutoLineDigitizer Gradio API for HuggingFace Spaces