numerical stability in sum-product BP

[dahtah]

Hi all, there seems to be an issue with BP in graphs with large number of labels. I've modified doSumProdInference.cxx to define a Markov chain where both univariate and bivariate factors are constant, so that the model is just that each variable is drawn IID from the uniform distribution over the labels. Running BP on a model with three variables and not-too-many labels works fine. With around 350 labels BP fails and the marginals of the intermediate variables don't sum to 1 like they should.
Here's the code: https://pastebin.com/DJdqKgvy