Improving Vision-Language-Action Model with Online Reinforcement Learning