[FEA] UMAP Python API to accept kNN graph directly

[cjnolet]

Both t-SNE and UMAP compute KNN indices and distances in the first stage of their computation and it is the kNN stage that needs to support different distance measures and metrics.

There are several reasons why being able to accept the kNN distances and indices is beneficial:

1. It allows the user to compute new distance measures that we may not necessarily support right away.
2. It gives the user an option of using other kNN algorithms, such as approximate variants
3. While FAISS is working on integrating most of the distance measures found in scipy to the brute force kNN, which is dense, this allows users to construct a KNN externally based on their own distance metrics, use a CPU-based KNN, or use any other GPU-accelerated KNN.
4. While we don't currently have a Sparse kNN option in cuML or on RAPIDS, this will enable sparse support in UMAP but giving users the option to use the outputs from libraries like SnapML (https://ibmsoe.github.io/snap-ml-doc/v1.4.0/knndoc.html) which accept sparse inputs.
5. This allows visualization tools to compute the kNN once for a specific metric and n_neighbors setting, only requiring the embeddings to be re-optimized when other parameters change. This offers a significant performance boost.
6. For the same reason as fix notebook bugs and add running results #5, this allows the user to compute the kNN once for a specific metric and n_neighbors setting, affording its use in parameter optimization or allowing the user to build many different layouts with the same NearestNeighbors inputs.

There is also a project on pypi called “umap-learn-modified” (https://pypi.org/project/umap-learn-modified/) that we might be able to use for API inspiration: https://github.com/lilab-bcb/umap/blob/master/umap/umap_.py#L1316.

This should mostly require python changes, as the input would just be bypassing the building of the kNN on the c++ side. It’s important that if fit() takes a kNN, the transform() also take a kNN, and assume the same metric was used.