DiffSensei: Bridging Multi-Modal LLMs and Diffusion Models for Customized Manga Generation