Resolve "Predictions for MPC"

Hi @nik20652 and @lis07128 I made a sketch of training with preditions. You can see Issue #47 (closed) for a more detailed changelog, but at a high level

• Refactor in trainer
  • Cleaned up some (seemingly) dead code, please check I didn't remove anything that is planned to be used (e.g. ModelParams)
  • Pulled dataset into the init fn and tried to reduce redundant state.
• Split out dataloader and dataset, added a dataset for predictions

Right now, we can switch btwn predictions and reconstruction by just setting the purpose string in trainer and pred_offset in dataloader. Please review the code to make sure it's:

• Legible, consistent with your guys style / architecture for the repo
• Doesn't remove any code I thought was dead that has plans
• Add any comments or suggestions in changes

When done, give a thumbs up and I'll test a full training run (just used test_training now) then merge.