Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition