Generalization Guarantees for Representation Learning via Data-Dependent Gaussian Mixture Priors