Offline Imitation Learning with Model-based Reverse Augmentation