BaseAgent, scalar losses

[jakegrigsby]

An update that makes it easier to customize agents in external projects

• creates a BaseAgent
• refactors loss computation to be less specific to (actor_loss, critic_loss)
• registration system (@register_agent etc) makes it easier to plug custom objects into example training scripts

And adds some RL things:

• n_step returns
• hl-gauss style value regression labels (instead of two-hot)
• extra options for normalization in MLP components