TalkingMachines: FaceTime-Style Video via Autoregressive Diffusion Models