Focusing on Language: Revealing and Exploiting Language Attention Heads in Multilingual Large Language Models