See #789.
According to @mblondel the callable kernel just works as a precomputed one.
Instead, it should be able to compute only those kernel values that are needed for the optimization (good luck with that ;).

Also pickling of lambda functions is no good.