Support the Manipulation of PyMC4 Models

[brandonwillard]

After the introduction of TensorFlow backend support (i.e. #4 ), we can implement functions to convert to-and-from PyMC4 models.

[brandonwillard]

FYI: This issue is more specifically about

1. the ability to obtain a usable "sample-space" TensorFlow graph (i.e. one that represents all the relations between random variables/terms) from a PyMC4 model object, and then
2. the ability to convert such a TensorFlow graph into a PyMC4 model object.

As of now, the only way to do 1. is to create the corresponding TFP objects and use the graph from .sample(). Requirement 2. isn't even remotely possible to do right now, because the output of .sample() only loosely corresponds to the TFP objects/classes from which it's generated, and, even after converting the TF sample-space graph to TFP objects, there isn't a clear way to construct a corresponding PyMC4 model.

Don't get me wrong, it's do-able, just a lot of painstaking 3-way patching between regularly updated/changing libraries.