Update to more accurate way of calculating ancestors

[ColCarroll]

Fixes #3165. There was a theano function I should have been using the whole time (theano.gof.graph.ancestors instead of theano.gof.graph.inputs).

I suspect accurate listing of parents may help with doing sample_prior_predictive (#3210) as well.

Both tricky cases mentioned in the issue are fixed, and here's another extra tricky one that has a nice graph. I am also rechecking the radon notebook now.

with pm.Model() as model:
    a = pm.Normal('a')
    w = pm.Normal('w', a)
    x = pm.Deterministic('x', w)
    y = pm.Deterministic('y', x + a)
    u = pm.Deterministic('u', y + w)
    z = pm.Normal('z', u + y, x)
pm.model_to_graphviz(model)

[ColCarroll]

Radon notebook graphs all look identical.

[twiecki]

Not sure why travis is complaining, doesn't seem like any tests weren't passing.

There are, however, many warnings (unrelated to this), like:

/home/travis/build/pymc-devs/pymc3/pymc3/distributions/distribution.py:424: DeprecationWarning: The truth value of an empty array is ambiguous. Returning False, but in future this will result in an error. Use `array.size > 0` to check that an array is not empty.
  shape = tuple(shape or ())

[ColCarroll]

Yeah - I think the new numpy is emitting a lot more warnings. I can tackle some of those in a new PR. This might be related that we already fixed: arviz-devs/arviz#271

The job exceeded the maximum log length, and has been terminated.

[ColCarroll]

Is this ok to merge? #3214 may supersede it, but this still fixes some issues.