Mechanistic Understanding and Mitigation of Language Confusion in English-Centric Large Language Models