Fix render_model for intervened variables

[SamWitty]

Currently, applying pyro.render_model to a model after an intervention removes edges out of the intervened variable, rather than into the intervened variable.

Pre Intervention Example

Post Intervention Example

[eb8680]

This is (currently) expected behavior from Pyro's point of view - intervene and do have a SWIG-like default semantics, in which intervened variables like training disappear because they can be marginalized away if not observed, not because of the semantics of intervention itself.

To get the behavior proposed here (still a worthwhile goal since many people will have the same question), we could do the following:

1. use an effect handler to change the meaning of do and intervene so that the training random variable is removed entirely (see e.g. this example Interventional handler in the counterfactual docstring)
2. enable rendering of pyro.deterministic sites in Pyro - see Investigate and possibly fix upstream Pyro issues with rendering deterministic dependencies #41.