Benchmark should include "normal" input data

[hbf]

Currently, the benchmark on the project front page uses points coming from a sphere, which is a hard case for the algorithm. The performance on "normal" point sets is much better, which is not reflected in the chart.