Experimental MultiGraph support

[Lamaun]

See #62

[Lamaun]

Should I write some tests for this? Or is something else keeping this from a merge?

[dschult]

Sorry this is not getting much attention.
Here's a review:

• I'm not sure this change will work with the METIS adjacency_list data structure. But I assume you have checked that it works. Can you add some tests -- they will both check that something doesn't mess it up later, and serve as a very basic form of documentation of how it should work.

• dict(G[u])[v] should be simplified to G[u][v] ... or you could make the inner loop:

for v, keydict in G[u].items():
    adjncy.extend(index[v] for key in keydict)

- docstrings for private functions don't matter much, but could you change _convert_multi_graph and _convert_graph to docstrings like:

    """Convert a Graph to the numbered adjacency list expected by METIS.
    """

That way they are a single line and under 80 chars so Sphinx and pycodestyle are happy.

- To allow subclasses of networkx graphs, replace ```if(isinstance(G,nx.MultiGraph)):``` with ```if G.is_multigraph()):``` 

Thanks!