TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models