Enhancing Diffusion Models for Inverse Problems with Covariance-Aware Posterior Sampling