[FEA] Batching NN Descent

[jinsolp]

Description

This PR implements batching NN Descent. It will be helpful for reducing device memory usages for large datasets (specifically if the dataset is kept on host).
index_params now has...

• n_clusters: number of clusters to make. Larger clusters reduce device memory usage. Default is 1, in which case it doesn't do the batched NND.

Notes

• The batching approach may have duplicate indices in the knn graph (in rare cases) because sometimes distances calculated for the same pair may be slightly different. This results in putting the same index far apart after sorting by distances, making it difficult to get unique indices (which is done by looking at 2 indices before the current one).
  • handled by adding a max_duplicates for check_unique_indices in tests

Benchmarks

• Dataset for NND (no batch and batch) is on host
• Dataset for brute force knn is on device (but still won't be able to run with large datasets even if the data is put on the host because it brings the entire dataset to device anyway)
• The dataset is just a slice of the wiki-all dataset (88M, 768) to test for different sizes

[cjnolet]

So sorry for being so late to review this @jinsolp. I think this is very close, but in order for this to be useful to users, we should make it as easy as possible to use and document very clearly what the algorithms does (and why a user would want to use it).

[cjnolet]

/merge