add run_diversity, return hiddens from llm

[davidbrandfonbrener]

Added run_diversity.py to run search with actions selected to maximize distance from embeddings of previously seen actions.

This required modifying llm.py to add an option to return hiddens when generating (False by default) and montecarlo.py to store the global features in the montecarlo object.

Also fixed a typo I noticed in llm.py where "prompt" was misspelled.

[namin]

Thanks, this looks like a great start. I'll merged it in now, and play with it a bit later.