SUGAR: Efficient Subgraph-level Training via Resource-aware Graph Partitioning