Integrate metric logger into grpo training loop

[DNXie]

Integrate metric logger into grpo training loop with some basic logging

[felipemello1]

You probably want to make sure that the input is a dictionary, like here: https://github.com/pytorch/torchtune/blob/67ab86b94de9e7ac7dd9850113ebe69e2bbd307c/recipes/full_finetune_distributed.py#L909

I think that we will have an abundance of metrics coming from dataset and reliability metrics. This is how i envisioned it being used: https://fb.workplace.com/groups/1189731669410969/permalink/1279384097112392/

I understand that its a a simple PR just to get logging started. Just sharing where I think we should land after a few iterations.

[joecummings]

Yeah great callout, but lets land this one for now and iterate to that point.

[DNXie]

@felipemello1 Thanks for the suggestions! This is a simple PR to simply integrate the logger. In my next PR, I will add logging for various metrics using log_dict.
@joecummings Thanks for approval. I will go ahead and land this one for now.