Tensorflow fixes - #20

Some fixes to tensorflow imports and logging

• TF import was moved to top of main script, after arg parsing, to ensure environment variables are set before TF initializes. This avoids repeated TF imports.
• In order to avoid TF being loaded early via keras in model_loader.py, using lazy import in ResidualModelBase
• TF C++ error messages are suppressed in --cpu-only mode (since these are almost always CUDA / other GPU messages, and any other errors are expected to give a Python level error)
• The uv.lock was updated with regards to the latest TF dependency changes