Demystifying Multilingual Chain-of-Thought in Process Reward Modeling