model_to_graphviz display bug when missing value is involve.

[junpenglao]

x = np.random.randn(10) + 10.
x = np.concatenate([x, [np.nan], [np.nan]])
x = np.ma.masked_array(x, np.isnan(x))

with pm.Model() as m:
    a = pm.Normal('a')
    pm.Normal('x', a, 1., observed=x)
    pm.model_to_graphviz(m)

gives:

which is technically correct as we define x_missing as a NoDistribution: https://github.com/pymc-devs/pymc3/blob/ee57bed9aa5f73dc48081dd3c82267dd4f0f2834/pymc3/model.py#L1672-L1677

the reason of doing so is that we fill in the missing value into the original data: https://github.com/pymc-devs/pymc3/blob/ee57bed9aa5f73dc48081dd3c82267dd4f0f2834/pymc3/model.py#L1683-L1684

and to not double count the log_prob, the data tensor need to be a FreeRV (so we sample from it) with no log_prob (as the log_prob already accounted for in x in this example).

I think the easiest way to fix is to overwrite the _repr_ here with the _repr_ from parent.

[junpenglao]

Actually, to get the graphviz display correctly there is a bit more complicated, as currently the logpt is not connected to any ancestor. we can put a logic in model_graph.py so it replace NoDistribution to the underlying parent_dist, but the graph link will still be incorrect...

[nicospinu]

Thank you @junpenglao

[ricardoV94]

@junpenglao What would be the correct graph here? Something like this?

`