NeMo RL + Gym integration: dynamic sampling

Use cases, pain points, and background
We want to only train on mixed-reward examples to improve training efficiency.

Description:
Implement dynamic sampling. As soon as a group finishes, we check it for mixed rewards and kick off another group if necessary until the entire batch is full.

Design:
Probably need to touch a few files, but most likely we need to shuffle some responsibility around in the rollout + Gym logic.

Need to figure out how to yield one prompt at a time from the train dataloader and see how that works. Need some investigation into torch StatefulDataloader and possibly the collation function.

May need to do one collation at the end.

Out of scope:

Acceptance Criteria:

• Implementation of the above + tests in NeMo RL.