Equivariance Everywhere All At Once: A Recipe for Graph Foundation Models is accepted to NeurIPS 2025