Initial notes (WIP):

• Metrics: these are importable (like VMRSE)
• Models: these are importable provided done with:

from the_well.benchmark import models
fno = models.FNO

• Data structure for models: the trainer rearranges, assumes time in channel dim, we can add attribute on emulator that says whether this is expected. See tutorial.
• Wrapping The Well: create a base class that has model attribute. This can be set at init. Subclasses can fix this. Similar to GP subclassing currently impled.
• Training: is handled in separate module. To use with fit, we can wrap a trainer in the Well base class.
• Autoregressive training: how/what is implemented here in the well?
  • No autoregressive training - next m time-step prediction
• Autoregressive inference: how/what is implemented here in the well?
  • A rollout model is implemented here with AR predictions in the form of a moving batch.