test_exact.test_shortest_path_affinity fails

[stanleyjs]

Hi,

This test seems to fail in a way that is dependent on the sampling of the test dataset. I have been tinkering with this for a little while trying to get it to fail in a reproducible way, but it seems to be entirely dependent on the indices the test grabs. I have also tried to make it stop failing by replicating the code that generates the distances in the test inside of the base.shortest_path function, but I haven't gotten a reliable repro of that yet either. It seems to (maybe?) be something to do with the transition from sklearn.utils.graph_shortest_path.graph_shortest_path to scipy.sparse.csgraph.shortest_path