Remasking Discrete Diffusion Models with Inference-Time Scaling