Fairface collides with current torch/torchvision CUDA kernels

[trifle]

We have torchvision unpinned in requirements.txt, giving us the most up-to-date version that works with host drivers (in my case, 470.103.01 / CUDA 11.4).

This enables cuda in torch, but collides with the installed version of fairface. To run that, a temp fix is to manually step back to an older version:

pip3 install torch==1.7.1+cu110 torchvision==0.8.2+cu110 -f https://download.pytorch.org/whl/torch_stable.html

But we should consider either vendoring fairface or pinning torch and torchvision.
Note: Pinning might collide with the newer oxnn torch-based models from insightface.