[Feature] Implement VQ-VAE

[alexcbb]

Feature detail

The VQ-VAE model permits to reduce the input frames into a discrete vocabulary. It thus reduces the entire compute time of the entire model.

What needs to be done ?

• [ ] Reimplement the original VQ-VAE with existing repos
• [ ] Evaluate the network on a toy dataset
• [ ] Evaluate the network on videos
• [ ] Prepare unit tests
• [ ] PR the new feature