Hindsight Generative Adversarial Imitation Learning