Quest: Add Benchmarks and Pretokenized JSON for training

[gkielian]

Benchmarks

This is going to be a huge contribution.

Currently the get_batch function currently performs fast inference and tests for validation loss.

Create a new get_batch specifically for benchmarking.

This would utilize a json file (pretokenized), and has "question:" "answer:" for fields.

Most of the time we'd suspect just a single multiple choice answer (e.g. MMLU Pro), or a set of correct answers, and we can see 1 or 0 if the network has the correct answer from top 1 logit top 2 logits , etc.