Solving Inverse Problems via Diffusion-Based Priors: An Approximation-Free Ensemble Sampling Approach