Combined metric + Butina Kernels

[scal444]

To avoid n^2 distance matrices, we could compute tanimoto distances on the fly when running Butina. However, butina can be run on generic data, so any use case we want to support has to be done on a case-by-case basis.

[scal444]

This is low priority for us now because other clustering methods exist for larger datasets, and because of the high per-implementation cost + inability to make it easily generic, but if anyone has important use cases and thinks Butina is critical for their workflows, comment here and we can reprioritize.