Advancing Multimodal Reasoning: From Optimized Cold Start to Staged Reinforcement Learning