In-depth: Google TurboQuant cuts LLM memory 6x, resets AI inference cost curve