Distill the Image to Nowhere: Inversion Knowledge Distillation for Multimodal Machine Translation