add nndescent algorithm

[brj0]

This adds support for nndescent, which is a C++ implementation of pynndescent.

Although single queries are slower compared to the original Python implementation, the build time (which includes the creation of a k-NN graph for the training data) on my machine shows a significant improvement. Here are the results of ann-benchmarks for fashion-mnist-784-euclidean_10_euclidean:

Build time (s) tradeoff

Recall	nndescent	pynndescent
0.9667	13.609
0.9693	14.184
0.9876	21.162
0.9877		81.0728
0.9997	42.560	111.5696
1.0000	51.341	125.3250

[erikbern]

Nice!