LLMs Don't Have to Generate One Token at a Time: How Medusa and Multi-Token Prediction Cheat Autoregression