Enhancing Non-English Capabilities of English-Centric Large Language Models through Deep Supervision Fine-Tuning