Training-Free Self-Correction for Multimodal Masked Diffusion Models