Class for storing rollout information for RL trainers.
This function is used by the training engine to temporarily (after one interrupted rollout in decentralized distributed settings, without arguments) or permanently (for a training stage with shorter horizon, with arguments) narrow the step dimension in the storage.
The reverse operation,
unnarrow, is automatically called by
after_update (without arguments) or when the rollout length
varies in the training pipeline (with arguments).
See doc string for the