Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers