Learning from Graphs with Heterophily: Progress and Future