Multi-Source Diffusion Models for Simultaneous Music Generation and Separation