Diffused Heads: Diffusion Models Beat GANs on Talking-Face Generation