Preserve Your Own Correlation: A Noise Prior for Video Diffusion Models