RL Bug/Issue Tracker

This issue is a running tracker of all RL bugs/issues we identify:

- [ ] Logging for metrics returned by loss functions via `aux` is at a `self._iter_steps` level, whereas other metrics are logged at a `self.global_steps` level.
- [ ] The progress bar for GRPO/PPO is at a `self._iter_steps` level. We should possibly unify this with how we log metrics. It's a bit confusing for the user to pass `max_steps` at a `self._iter_steps`, and not at `self.global_steps` level.