Towards Balanced Continual Multi-Modal Learning in Human Pose Estimation