Speculative Decoding Pipelines: Draft-and-Verify for Production LLMs