MIM-Reasoner: Learning with Theoretical Guarantees for Multiplex Influence Maximization