Google’s TurboQuant AI-compression algorithm can reduce LLM memory usage by 6x google Article 345 words 44 votes