Accelerating LLM Inference with Lossless Speculative Decoding Algorithms for Heterogeneous Vocabularies