Google says its TurboQuant algorithm can diminish the memory capacity required for operating large language models by a factor of six, significantly lowering the financial burden of AI training