Speed up KMeans using FAISS

[lukegre]

I know that clustering probably isn't the biggest bottleneck in the process, but I've used FAISS in the past, which is a lot faster.
This provides a few more details: https://www.kdnuggets.com/2021/01/k-means-faster-lower-error-scikit-learn.html
This might make searching for the number of optimal clusters quite a bit quicker (20 x if the article is anything to go by).

[ArcticSnow]

Thank you @lukegre for the suggestion. This looks interesting indeed! As you said the biggest bottleneck isn't quite the clustering step. If you feel like it you are welcome to push a PR adding this method.

[lukegre]

Did a quick test, FAISS is much faster than regular Kmeans, but not than MiniBatch. So, probably not worth the effort.

[ArcticSnow]

Good. As you had pointed out, clustering is not the bottleneck. But thanks for suggesting and testing.