Compute testing data log-likelihood in addition to training data log-likelihood.

[wangkuiyi]

Currently we compute training data log-likelihood in every training iteration. This helps us plot the log-likelihood v.s. iteration curve and indicate the convergence of a training job.

We need in addition to this is compute testing data log-likelihood after the model get converged. This value indicate how "good" the model can explain new data, and will be used in model selection (learning the optimal number of states.)