Same Content, Different Answers: Cross-Modal Inconsistency in MLLMs