Dual RL: Unification and New Methods for Reinforcement and Imitation Learning