Airflow Docker build installs heavy NVIDIA CUDA dependencies via docling

Bug

When running the Docker build for this repository, the image build process pulls several gigabytes of NVIDIA CUDA dependencies.

Root Cause

The project includes docling, which relies on torch and torchvision. When uv builds inside a Linux Docker container, it pulls the standard PyPI Linux wheels, which default to bundling full CUDA binaries. Running uv add docling locally on a Mac or Windows machine bypasses this because PyPI delivers native CPU/MPS wheels for those operating systems.

Suggested Fix

We should force the build to look at the PyTorch CPU index.