A Self-Supervised Paradigm for Heterophilic Graph Representation Learning