Refactor evaluation code

[grantgasser]

• Clean up evaluation.ipynb notebook
  • plug and play essentially (e.g. avoid thing where GT and EST sources getting mixed up/backwards)
  • make it easy for people to evaluate their results
• Refactor plot_runs.py
  • so it stores data automatically when experiments are running
  • makes no assumptions (e.g. line assuming validation episodes every 5 episodes)
• Anything else that makes it easier to evaluate the trained model

[grantgasser]

#11 is a natural follow-up to this