Navigating the OverKill in Large Language Models