Implement data estimated initialization points

[jeffreypullin]

Currently, we implement initialization points that are basically the same as the prior. These were taken from one of Bob Carpenter's posts on the Stan forums I think - I don't have a link handy.

The other option would be to implement the data driven initialization points from the original Dawid-Skene paper. Having not implemented them, I don't know if they would make an improvement. They were probably more important in the context of the EM algorithm, but they may help in cases where the data and priors don't match well.