Computation of the KL divergence

[Hebbalali]

Dear Hugh,
When looking into the code of the computation of the KL divergence in the SVGP layer, i noticed that it computes the KL between q(u) and the prior p(u) with the mean of p(u) taken equal to zero. Which may not be adequate since for example in the inner layers a mean function XW is considered for the inner layers.

[hughsalimbeni]

The q(u) distribution doesn't include the mean function. The only difference would be that q_mu would have model the mean, but this isn't necessary as it's just a reparameterization. This follows the convention in gpflow.