GSINA: Improving Subgraph Extraction for Graph Invariant Learning via Graph Sinkhorn Attention