- don't use RNG for "task size" parameters (N, K, dims)
- add "good" kmeans test data (without singularities: K > unique points)