Never Lost in the Middle: Improving Large Language Models via Attention Strengthening Question Answering