baseline benchmark: add some clustering to results before reporting

[jan--f]

Right now the baseline benchmark reports the average OSD throughput and OSDs that deviate from this average by a certain percentage.
When two (or more) class of OSDs are in a cluster (think one class on spinners, another only on SSDs) and given the right amount of instances of each class this would report all OSDs as outliers. The average is a bad statistical tool here.

Implement some clustering after the test was run and report each cluster (hopefully OSDs of the same class) individually. With some tuning the clustering could be quite good and make the output much more informative on heterogeneous clusters.