MJ-BENCH: Is Your Multimodal Reward Model Really a Good Judge for Text-to-Image Generation?