Hierarchical GPs in GPyTorch

[GStechschulte]

I have implemented a hierarchical GP with the following components:

• Custom mean module that learns a ConstantMean for each group.
• Covariance module that is a product of an RBF kernel and an Index kernel where the RBF kernel models the continuous dimensions and the Index kernel models group-specific effects.

In this Gist, you will find a reproducible example.

The model architecture is inspired by the multi task GPs examples, but is different to account for the subtleties in modeling data with a group-like structure, e.g., varying group sample sizes.

I noticed there are some open issues (#996) and discussions scattered throughout forums regarding hierarchical modeling in GPyTorch, so if the maintainers think it would be a nice addition to the multi task docs, I would be more than happy to open a PR.

Feedback or questions is welcome :)