The RL loop should print or log a few example prompt/response/targets

[allenwang28]

We're using the reward/loss as our proxy for RL correctness, which is understandable. But for understanding whether or not the RL loop isn't reward hacking, it would be helpful to additionally log a few of the prompt/responses/targets per step (?) in wandb and/or logger.

[felipemello1]

@DNXie is looking into creating the table and print it. We will log once the logger is ready.