[WIP] RL goes brrr

[lewtun]

Adds support to run GRPO with vllm in multi-node settings. Needs huggingface/trl#3094 to be merged first

TODO

• Add doc
• Validate we can train multi-node with 7B model
• Pin trl to new release

[lewtun]

Thanks for the fixes! LGTM with some questions about the settings of some args.

Could you also update the README as running the GRPO training now requires one to spin up the server on e.g. 1 device and then running training on the remaining 7