Can message-passing GNN approximate triangular factorizations of sparse matrices?