On kernel-based statistical learning in the mean field limit