Efficient Active Imitation Learning with Random Network Distillation