Deep kernel learning example: add mini-batching

[st--]

This should prevent the overfitting issue, as discussed by Ober et al.

[willtebbutt]

Are we able to do that in this repo? We'd need a pseudo-point approximation that permits minibatching in order to do that, so probably we'd have to do it in ApproximateGPs.

[st--]

Ah good point :|

[st--]

Maybe an alternative would be just to show train/test curves to demonstrate that you need to do something.

[willtebbutt]

Makes sense to me 🤷