Better PAM initialization with BUILD

The PAM/k-medoids implementation appears to implement SWAP, but not the BUILD part for initializing PAM. Instead you have to provide good starting medoids.

I tired benchmarking it on a larger data set (well-known 20news data set in the `sklearn.datasets.fetch_20newsgroups_vectorized` version; with cosine distance and k=20) and the runtime of pyclustering kmedoids was extremely high (2000 sec), likely because of the poor random initialization.

With BUILD from the `kmedoids` package, I can reduce the run time to 338 sec. Nevertheless, `pam` from `kmedoids` is just 37 seconds, including 4.6 seconds for BUILD. The `fasterpam` variant finishes in 336 msec with random initialization.

It would also be good to get access to the final loss after optimization as well as the number of iterations. Then I could check if it ran into the maximum iteration limit, and if at least the result quality is comparable.