Add CLARA, FastCLARA, FasterCLARA

[kno10]

CLARA roughly does:

• subsample the data
• run PAM (FastCLARA: FastPAM, FasterCLARA: FasterPAM) on the sample
• compute the total deviation on the entire data set for these medoids
• return the best result found with multiple subsamples

This may seem like a trivial addition at first (and it would indeed only be a few lines in the Python wrapper) BUT:

• this package currently does not include any distance functions, but operates on precomputed distance matrixes only
• if you already have the distance matrix, just use FasterPAM and you will be fine
• a meaningful implementation of these only computes the distance matrix on the subsample - which needs a data matrix as input and distance functions
• for many users it will still be more convenient to handle the subset/sample within their own application

Hence a rough implementation plan would be

• design an API for computing distances compatible with typical users (python wrapper, rust native users)
• implement a decent choice of distance functions
• implement CLARA
• tests
• update the Python wrapper

Adding distance function will also be necessary for CLARANS #6 BanditPAM #2 or coreset approaches #4