Seeing Eye to AI: Human Alignment via Gaze-Based Response Rewards for Large Language Models