On the Stability and Generalization of Learning with Kernel Activation Functions