A Divergence Minimization Perspective on Imitation Learning and State-Marginal Matching