Run more RL experiments. Wait less for GPUs.

Implementation for FP8/INT8 Rollout for RL training without performence drop.

Ongoing research training transformer models at scale