On Learning Multi-Modal Forgery Representation for Diffusion Generated Video Detection