Two Heads Are Better Than One: Boosting Graph Sparse Training via Semantic and Topological Awareness