report rl and optimization step differently

when we have off-policy steps in PPO the metric gets averaged across the different steps. This can be very hard to reason about, we should report each optimization step as train/optim_step and every rl step as train/rl_step to avoid this confusion.