Re-ranking Reasoning Context with Tree Search Makes Large Vision-Language Models Stronger