Selective Timestep Weighting and Advantage-Based Replay for Sample-Efficient Diffusion RLHF