Deep Deterministic Policy Gradient implementation.

This assumes that the Actor network will output {nil, num_actions} actions, and that the Critic network accepts the "actions" input with the same shape.

Actions are deemed to be in a continuous space of type :f32.