Witnessing Adversarial Training in Reproducing Kernel Hilbert Spaces