On the Role of Attention Heads in Large Language Model Safety