MS-ILLM models and training scripts

[mmuckley]

This pull request implements, the generative image compression model from the following paper:

MJ Muckley, A El-Nouby, K Ullrich, H Jegou, J Verbeek. Improving Statistical Fidelity for Neural Image Compression with Implicit Local Likelihood Models. In ICML, 2023.

The PR includes:

• A recursive bottleneck structure for implementing VQ-VAEs.
• A new VQ-VAE model, based on the VQ-GAN paper, that has XCiT attention for improved memory efficiency.
• A modification of the OASIS U-Net discriminator that clips the discriminator in the upsampling path corresponding to the VQ-VAE latent space.
• A PyTorch Lightning module for HiFiC-style training of hyperprior models for targeted bitrates.
• A PyTorch Lightning module for fine-tuning the HiFiC-autoencoders with MS-ILLM

Installation instructions and basic usage are covered in the README in the illm project folder