Refactor affinity module

[pavlin-policar]

Description of changes

Adding this kind of functionality complicated things quite a bit, so a pretty major refactor was necessary to keep things manageable. Notable changes include:

• build_knn_index no longer computes the neighbors, but only returns the correct index. This lets us pass in already existing KNNIndex objects to affinities.
• All Affinity classes can now be called in two ways. 1) with a data matrix, as is standard, or 2) using an existing KNNIndex object. This is in preparation for scanpy integration, where we will have to use existing KNNGs.
• KNNIndex objects now have the default parameter `metric="euclidean". This just makes it easier to debug or instantiate manually.
• Add nearest_neighbors.PrecomputedNeighbors which allows us to wrap the standard KNNG representations.

Because there are now a whole bunch of possible ways to call this API, I added a test module which runs over the different ways to create an iris embedding. See tests/test_different_usages.py.

Includes

• [X] Code changes
• [X] Tests
• [ ] Documentation

[dkobak]

Wow, this precomputed thing escalated pretty quickly :)

[pavlin-policar]

Haha, yeah. Things got pretty complicated when I just patched things up. So I tried to simplify it, but all the different ways we can call things now and string them together have made things a bit messy and necessarily complicated.

I'm trusting that my tests are thorough enough that nothing broke here. Most of the major things are tested pretty well though.