Experiment: Rollout #252
Description
Once we have a model with a good loss function, we should do rollout training.
The models generally look ok with inference for 2-3 time steps, but after that artifacts show up which "explode" over time. My hunch is that rollout will remove these artifacts to some extent.