Add support for GPT-JT model + make it easier to plug in other LMs

[anas-awadalla]

GPT-JT seems like a great choice for the LM side as it outperforms many other open-source models and is considerably smaller. Currently, we only support OPT models, but in general, it would be great to add more options.

The current design of the Flamingo class makes it difficult to plug in any huggingface model, as you would need to write a class specifically for each model family. We should explore more modular ways to build this.

[i-gao]

#66