Scalable Gromov-Wasserstein Learning for Graph Partitioning and Matching