How TurboQuant uses a random rotation to precompute its quantizer, and why skipping the training step changes the operational story.