Introduce text modality into the repo

[bhavul]

Currently the repo is focused on control tasks. This is an issue so we can work on including pure text based task and use autoregressive pre-training on the same.

This likely means :

• Extending the code to work with text (TextTokenizer, Prediction, etc) [Note GPT-2 is already existing so we don't have to write the model]
• Using the current workflow of the code alongwith conventions
• Make the code compatible to train both control and text tasks together by merging the data batches

As for how this could be done, I wrote a doc couple of weeks back explaining different parts of the repo and how and what are the places where code changes are to be made : https://docs.google.com/document/d/16rNVqRucDMz4yqiYJz8NZFPHMGvzO0emb0KkxjAlfXo/edit#heading=h.huk7442i8c1n

[bhavul]

The code has been merged for addition of text modality now, closing the issue. Feel free to reopen if needed.