On kernel-based statistical learning theory in the mean field limit