Deconfounded Imitation Learning