Add pre-trained models from HuggingFace and add accuracy testing

[darsnack]

This PR builds on the great work by @Alexander-Barth in FluxML/MetalheadWeights#2 to add pre-trained weights for VGG and ResNet from torchvision.

We ran into issues with Github LFS limits on MetalheadWeights. Instead, we are using HuggingFace to host our models now which is really ideal for our use case.

This PR adds:

• Pre-trained models for VGG (w/o BN) + ResNet
• Updated quickstart guide
• Updated contributor docs for how to others can add weights via HF
• Refactored tests to check the accuracy of pre-trained models (by just checking against a single reference image from ImageNet)

Close #86, closes #72, and closes #143.