Add support for pre-tokenized streaming dataset finetuning

[boomanaiden154]

This PR adds in support for performing fine-tuning training with a streaming dataset that has already been pre-tokenized.
The test passing is dependent on the fix in #600.

[dakinggg]

Thanks! Will take a look after we resolve #600

[alextrott16]

I am not opposed to supporting pre-tokenized streaming datasets. But, I am concerned that the tokens/labels structure used to identify pre-tokenized formats is too arbitrary. It's an issue mostly because we don't have any tools that someone can use to build a streaming dataset with the tokens/labels structure.

Rather than having a "backdoor" in StreamingFinetuningDataset that recognizes pre-tokenized formats from the tokens/labels structure, I'd prefer that we keep with the prompt/response structure and simply allow the code to determine whether tokenization is required based on whether the prompt/response values are strings or bytes. I think this will make it easier to simply add a pre-tokenization option to our existing tooling.

Please let me know if this ask is unclear. Thanks for helping to grow our codebase!

[boomanaiden154]

That's a pretty reasonable request, especially given there's nothing upstream using this! I'll change the implementation to follow your suggestion.

I'm planning on writing up some tooling eventually designed for upstream use that uses this format (probably as a part of #611), but that'll come later.

[alextrott16]

Thanks for adding this and incorporating the suggestions! LGTM!!

[dakinggg]

This has been done by #945