SFT matches RL if you MCMC the training data first