v0.2

Significant performance improvements. Motivated by and ideas from commit amb@fd99396, thanks!

• Using sparse eigenvector computation by default now. Old dense method can be chosen as fallback with a property
• Much better implementation of affinity matrix scaling