On the approximation capability of GNNs in node classification/regression tasks