Implement student-t likelihood

The projection for the student-t likelihood would be a nice addition to the package. The student-t is not a log-concave likelihood, and thus the implementation using the same framework as for other likelihoods is likely to cause some problems. However, there is another idea of doing the projection similarly as for the Gaussian case by expressing the t-distribution as a scale mixture of Gaussians, and then using the Gaussian projection with the observations weighted by their precisions. This is not the same as measuring the KL-divergence between the t-distributions but should be easy to implement and might give reasonable results.

Start by locating the existing Gaussian projection and the likelihood-projection entry point. Compare the proposed scale-mixture approach with the current framework, then determine how the student-t behavior should be validated and what results would count as reasonable.