An Empirical Study on Cross-lingual Vocabulary Adaptation for Efficient Generative LLM Inference