Hetero$^2$Net: Heterophily-aware Representation Learning on Heterogenerous Graphs