Integrate graph objects into hdbscan

[JanRhoKa]

Adapt the HDBSCAN function for graphs.

Create a new metric called "graph", which run HDBSCAN on graphs using an adjacency matrix of the graph, in scipy.sparse.csr_matrix format.

Modified the _hdbscan_sparse_distance_matrix function for this, as no graph has to be calculated from the data. Additionally an example of the function is provided under examples\plot_hdbscan_graph.py which shows the communities calculated by HDBSCAN, the communities calculated using the precomputed metric and the time save by integrating the graph directly into the function.

[ivirshup]

Glad to see this PR opened @JanRhoKa!

I've made a few comments about code organization.

My main thinking here is that the whole idea of this PR is to allow users to weight the graphs by means other than mutual reachability of a distance matrix. I think this could be done with smaller changes to the code in a few places.

For example, the changes I suggested to the test is more "If a user passed mutual reachability weights, the result should be equivalent to passing distances and letting hdbscan compute mutual reachability weights." Right now it looks like you're testing that the result of using mutual reachability weights vs. just the distances directly is minimal.

---

Minor other point: I would see if you could revert the unrelated formatting changes.

[lmcinnes]

Thanks for the updates. This looks like it is starting to come together. The test failures are due to importing networkx in the test module; but it doesn't seem to be actually used there. Perhaps you can remove the import.

It might also be useful (to help ensure the feature sees usage) to include some documentation for this. Perhaps a short tutorial on clustering a graph with hdbscan? Perhaps based on the example you already have?