MMGenBench: Evaluating the Limits of LMMs from the Text-to-Image Generation Perspective