Show HN: I wrote an LLM inference engine in pure Go – 48 tok/s zero dependencies