Provably Breaking the Quadratic Error Compounding Barrier in Imitation Learning, Optimally