Add DirichletProcess random variable

[dustinvtran]

A Dirichlet process random variable is added. It can take arbitrary tensor dimensions for the concentration parameter and base distribution.

The only problem I have with DirichletProcess is how to perform practical inferences. Because all randomness is hidden inside a while loop, the only inferences we can do are ones that forward sample. More work needs to be done that will allow, e.g., replacing the Beta priors with Beta variational approximations within the while loop. This requires additional features possibly in the copy() function.

todo

• Generalize the sample to allow non-scalar alphas.
• update model compositionality and iclr2017 pages
• add unit test