Add lola

[newtonkwan]

Adds LOLA to the set of strategies.
Goal: LOLA plays against LOLA and shapes the opponent to learn.

• add argument agent_states to PPO, PPO_gru, DQN update() function
• Add offline actor critic naive learner with policy gradient (Foerster 2017) with simple replay experience buffer. EDIT: Our NL uses advantage instead of a baseline.
• Naive learner learns to defect against on another, which reproduces the findings of the LOLA paper
• Add LOLA-DiCE implementation
• pull in new runner and implement refactored lola.
• Get LOLA to shape NL.