DreamPRM: Domain-Reweighted Process Reward Model for Multimodal Reasoning