VOCABTRIM: Vocabulary Pruning for Efficient Speculative Decoding in LLMs