From slow bidirectional to fast autoregressive video diffusion models