RAPHAEL: Text-to-Image Generation via Large Mixture of Diffusion Paths