In-Process Global Interpretation for Graph Learning via Distribution Matching