DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation