Fast Training of Diffusion Models with Masked Transformers