Supervised and discrete metric support

[dillondaudert]

v0.2 added support for multiple "views" of the input data, each with their own distance metrics.
However, for categorical or discrete measures, some additional care is needed when constructing and combining
the fuzzy simplicial sets.

See the conversation in the Python umap issue:

• computational details for supervised/semi-supervised dimension reduction (umap#624)

Related:

• Arbitrary data support #27
• WIP: supervised learning, data views #36