In-context Learning as Maintaining Coherency: A Study of On-the-fly Machine Translation Using Large Language Models