Functional Regularization for Representation Learning: A Unified Theoretical Perspective