Provably Efficient Adversarial Imitation Learning with Unknown Transitions