Fix inconsistencies with scoring ties in networkx

[bannanc]

This is a list of my planned fix for the problem in PR #68. First, It appears that the test that is failing does in fact have a more complex set of molecules so it makes a little more sense why one is failing. Essentially what this comes down to though is what happens when there is a tie in the scoring function in the find_pair function in the cluster_graph code. I had written SMIRKS based on my understanding where that function would pair atoms to storage places in the order they were given when there were ties. The problem in PR #68 appears to be that in 3.5 this was the case, but there was a change in networkx that leads to the ties being treated slightly differently.

Ultimately this shouldn't be behavior that depends on networkx. I think it would be preferable if the score was updated based on neighboring atoms when there is a tie on the current atoms...

I'm still thinking about how to generalize this, what if you really have two identical molecules or very symmetric molecules where there is a REAL tie in the score? I will add more detail about how I found the problem in a little bit, but I wanted to do an initial brain dump.

@janash I'm getting closer to narrowing down the issue.

[bannanc]

I need to add more details about how this causes issues in the overlaying of fragments, but as a temporary fix, I'm going to remove the problematic test because it asks to overlay bonds that are REALLY different so its not a very practical test.

The problematic test is this one:

    # 9
    (make_cluster_graph(['C1CCCC1', 'C1=CNC=C1', 'CO'], layers=2),
     "[#6H1X3x2r5+0a,#6H2X4x2r5+0A,#8H1X2x0!r+0A:1](-;!@[#1H0X1x0!r+0A])-,:[#6H1X3x2r5+0a," \
     "#6H2X4x2r5+0A,#6H3X4x0!r+0A:2](-,:[#1H0X1x0!r+0A,#6H2X4x2r5+0A,#7H1X3x2r5+0a])-;!@[#1H0X1x0!r+0A]",
     "[#6H1X3r5x2a,#6H2X4r5x2A,#8!rH1X2x0A;+0:1](-;!@[#1H0X1x0!r+0A])-,:[#6!rH3X4x0A," \
     "#6H1X3r5x2a,#6H2X4r5x2A;+0:2](-,:[#1!rH0X1x0A,#6H2X4r5x2A,#7H1X3r5x2a;+0])-;!@[#1H0X1x0!r+0A]"
     ),

Which attempts to overlay the

• one of the C-C bond in cyclopentane
• one of the C=C bonds next to the N in pyrrole
• the C-O bond in methanol

I am going to put in a PR removing this test, switching to python3.6 and addressing issue #75

[bannanc]

I removed the one problematic test in PR #68, but otherwise this is a longer term issue for someone to consider.