Feat: spherical / normalized centroids

Hello! In our own benchmark, `fastkmeans` achieved an **impressive 3x speedup** compared to the kmean of `faiss-gpu` that we used previously. Thanks for such a brilliant project!

In Faiss, when using `spherical=True` during Kmeans clustering, the centroids are normalized after each iteration. This will helps a lot when the original data metric is based on angular(Cosine or Dot) instead of L2.

https://github.com/facebookresearch/faiss/wiki/Faiss-building-blocks:-clustering,-PCA,-quantization#additional-options

I am curious if this feature could be applied in this amazing library.

Start at the library's K-means entry point and inspect how centroids are updated between iterations. Compare the requested spherical behavior with Faiss's clustering option, then add coverage showing whether centroids are normalized after each iteration and document the resulting API behavior.

Written by the indexing model from the issue text.