Efficient Imitation Learning with Conservative World Models