What Matters for Adversarial Imitation Learning?