BF: fix inference example reliance on node order

[matthew-brett]

We were getting not-deterministic errors from the inference algorithm example script in Python 3. I believe these were due to the following code in calc_marginals_sumproduct, where G is a graph:

    for node in G.nodes():
        potential = multiply_potentials(*[messages[(src,node)] for src in G.neighbors(node)])
        marginals[node] = normalize(potential)

    return marginals, potential.sum()

This code is iterating over the elements of an underlying dictionary (G.nodes()), so potential.sum() at the end returns some value effectively randomly chosen from all the nodes in the dictionary. I think that we actually want:

    for node in G.nodes():
        potential = multiply_potentials(*[messages[(src,node)] for src in G.neighbors(node)])
        marginals[node] = normalize(potential)
        potentials[node] = potential

    return marginals, potentials[target_node].sum()

where target_node = 'burglary'.