Advancing calibration for stochastic agent-based models in epidemiology with Stein variational inference and Gaussian process surrogates