Add RLHF example

[younesbelkada]

What does this PR do?

First of all thanks a lot for bringing this repository and making the LION optimizer more accessible
I did some quick experiments in the context of RLHF using Transformers Reinforcement Library (trl), and managed to get an interesting convergence using this optimizer (by diving the learning rate by 3 as suggested), and I wanted to share it here, and why not merge the example script on this repository and/or add few instructions about this in the README

Related huggingface/trl#152

[lucidrains]

@younesbelkada ah, thanks for the PR, but i feel that's a bit outside the scope

would welcome that you share a few training runs in the discussion though. we don't even know if this optimizer will pass the test of time

[younesbelkada]

Sure, no problem, here is the run I made: https://wandb.ai/distill-bloom/trl/runs/slgy199e?workspace=user-younesbelkada

And in comparison to adam: