MirrorBench: Evaluating Self-centric Intelligence in MLLMs by Introducing a Mirror