Using Random Noise Equivariantly to Boost Graph Neural Networks Universally