add to elbo term that is max when rotation angle is zero

[rrtucci]

Consider NbTrolsModel, for example. In this case, z is a vector of rotation angles \theta_j \in [0, 2\pi]. There are more of these angles than are necessary to fit a discrete prob distribution P(y, x) where x is the input and y the output. We are attempting to maximize ELBO(\lambda) with respect to a vector of parameters \lambda. So instead of ELBO, maximize

ELBO(\lambda) - \sum_j [ P(\theta_j|\lambda) - P(\theta_j=0|\lambda)]^2

or something like this that rewards making zero as many \theta_j as possible