Automate choice of posterior approximation in VI algorithms

[dustinvtran]

The last feature that we don't have for ADVI. But even better than ADVI in that we pieced together its fundamentals to benefit algorithms across the board: automated transformations, default posterior approximations, and different gradient estimators.

We could have automated the choice in ed.VariationalInference instead of specific classes like ed.KLqp and ed.KLpq. I chose not to because VI algs such as ed.BiGANInference and ed.WakeSleep always need to specify per-data point approximations.