Interpretable Imitation Learning with Dynamic Causal Relations