Steering Masked Discrete Diffusion Models via Discrete Denoising Posterior Prediction