Correcting What You Cannot See: Credit Assignment for Perception Distillation in Multimodal Reasoners