Rethinking the Diffusion Models for Missing Data Imputation: A Gradient Flow Perspective