Causal Semantic Communication for Digital Twins: A Generalizable Imitation Learning Approach