Old tokens requantized turbo3_tcq to turbo2_tcq. ~30% extra memory savings at acceptable quality cost for tokens with negligible attention weight