Dual parameterization for SVGP

[Turakar]

I currently do not have the time to understand this paper completely, but I want to raise your attention to it:

Dual Parameterization of Sparse Variational Gaussian Processes

On a first read, it seems like there's plenty of optimization potential in how we do variational inference in GPyTorch, both in accuracy and performance.

[gpleiss]

It's a good paper, and we'd definitely be interested in adding it if someone is willing to implement and put up a PR.