Refactor `ensemble.py` implementations using SavedModel

[dustinvtran]

@jereliu Currently the ensemble.py implementations in edward2/baselines compiles model graph by loading model hyper-parameters from flags, and then load model weights from checkpoints. This creates challenges for both model reproducibility (if there's a mismatch between the default and the actual model hyperparameters) and for code reuse.

A better approach is to use SavedModel, so the ensemble script doesn't have to depend at all on re-defining the TF model code, and we can create a general-purpose ensemble script that aggregates black-box model predictions.

https://github.com/google/edward2/issues/366