Add diffusion model implementation

[lesterphillip]

Hi, I'm adding a PR for a basic diffusion model implementation.

To test it, you can simply replace the configuration file in run.sh using the one in conf/diffusion.yaml

Let me know if you encounter any bugs or have any suggestions for revision.

[unilight]

Thanks Lester, really nice work! For now, I think we just need to unify some code structure and style. I'll try to test this model on other datasets (VCC2020 for instance) later.

[lesterphillip]

Thanks for checking the code! I've made some changes but here's a summary:

• I made the model signatures for both Taco2AR and Diffusion to be the same to remove most of the if conditions in the train and decode scripts.
• I transferred the feature normalization inside the model class (for both Taco2 and Diffusion) instead of the training loop.

Let me know if there are other changes needed!