Upstream fetch

…d properly before starting the fix...

…h. Two failures remain: TestNutsCheckTrace.test_bad_init raises a RuntimeError instead of a ValueError, but that seems to be the fault of parallel_sampling.py chaining the exception. Then test_step_exact for SMC gives different traces. Finally, sometimes, test_glm_link_func2 gets into a deadlock. Im not sure why. Will check in TravisCI if they pop out.

…dded `conditional_on` attribute of every distribution. This function does a breadth first search on the node's `logpt` or `transformed.logpt` graph, looking for named nodes which are different from the root node, or the node's transformed, and is also not a `TensorConstant` or `SharedVariable`. Each branch was searched until the first named node was found. This way, the parent conditionals of the `root` searched node, which were only one step away from it in the bayesian network were returned. However, this ran into a problem with `Mixture` classes. These add to the `logpt` graph, another `logpt` graph from the `comp_dists`. This leads to the problem that the `logpt`'s first level conditionals will also be seen as if they were first level conditional of the `root`. Furthermore, many copies of nodes done by the added `logpt` ended up being inserted into the computed `conditional_on`. This lead to a very strange error, in which loops appeared in the DAG, and depths started to be wrong. In particular, there were no depth 0 nodes. My view is that the explicit `conditional_on` attribute prevents problems like this one from happening, and so I left it as is, to discuss. Other changes done in this commit are that `test_exact_step` for the SMC uses `draw_values` on a hierarchy, and given that `draw_values`'s behavior changed in the hierarchy situations, the exact trace values must also be adjusted. Finally `test_bad_init` was changed to run on one core, this way the parallel exception chaining does not change the exception type.

…n distribution.py.

…ss of networkx.DiGraph

… is now stored in a networkx.DiGraph. Networkx is then used to compute subgraphs, and perform topological_sort on the graph during draw_values execution.

…e robust against theano autonamed ops.

… be done to get the correct behavior in python 2.

… This way it is easier to get the edge colors for networkx draw method. Added a toy draw function to show the basic draw features we can use from networkx.

…, even in the presence of plates of the same shape.