Feedback in Imitation Learning: The Three Regimes of Covariate Shift