Create synthetic MMLU via GPT-4

[nouhadziri]

Creation process:

1. Select random few-shot examples from each category in MMLU.
2. Prompt GPT4 with the few-shot examples to generate similar (but not identical) questions, options and responses.
3. Parse GPT4 outputs to extract responses and filter malformed outputs.
4. Upload data to HF.

[natolambert]

@nouhadziri this looks good. We need to move the location though to match other data scripts
scripts/data/...

[hamishivi]

Gonna close this since I think we've moved on -- feel free to reopen + prep for merging if you think otherwise!