Change to canonical form for NormalGamma and NormalWishart?

[kaselby]

Currently, the NormalGamma distrbution returns (mu, tau) when sampled. This distribution, however, is only used (as far as I know) as the conjugate prior to the NormalCanon, which needs the parameters in the canonical form (mu*tau, tau). This is also the case (I believe) for the NormalWishart and MvNormalCanon.
This is obviously not a massive issue as it can simply be converted when needed, so I understand if it is low priority. It is causing me some annoyances with typing and dispatch, however, since I cannot simply use the samples from the prior as the parameters for the likelihood (whereas with other distributions I can). What are your thoughts on changing this? I can submit a pull request with the changes if you are okay with it.