When Many-Shot Prompting Fails: An Empirical Study of LLM Code Translation