Multi-dimensional sufficient statistics

[felix-clark]

I haven't found much in the way of documentation or examples of this, but it should be possible to, for instance, use separate simultaneous predictors for mu and log(sigma^2) in a Gaussian response.