Advantage Weighted Matching: Aligning RL with Pretraining in Diffusion Models