Identifying and Solving Conditional Image Leakage in Image-to-Video Diffusion Model