gpu additions

[LouisFaure]

As discussed on issue #17, here is a pull request with gpu accelerated code added! A simple device="gpu" is needed when calling harmony.core.augmented_affinity_matrix and harmony.plot.force_directed_layout.

The code has been tested and needs rapids-0.17 as well as cupy>=9.0 (I haven't added these packages to the requirements, I instead mention these in the README), right now cupy 9.0 is in beta, but a stable version should be out in the coming months.

Here is what is happening when running on GPU:

• cuml.NearestNeighbors is used instead of sklearn.neighbors.NearestNeighbors and sc.pp.neighbors
• cuml.LinearRegression is used instead of sklearn.linear_model.LinearRegression
• cugraph.ForceAtlas2 is used instead of fa2.ForceAtlas2
• cuml.PCA is used instead of sklearn.decomposition.PCA

As a side note, _mnn_ka_distances function has been slightly modified for speedups.

Here is a comparison using the included example data:

[ManuSetty]

Fantastic - thank you very much! Appreciate your contributions.