Q-learners Can Provably Collude in the Iterated Prisoner's Dilemma