Improve benchmark scenarios

[torkleyy]

While the benchmarks can show a rough difference, they're very far from real-world and they favor some approaches more than others I think. One point which should be improved for example is the "rendering" in the pos_vel benchmarks, because there's no code executed. I think it would also be great to have a benchmark with more systems than two.

I hope we can collect some ideas here and work out detailed specifications what the benchmarks should do. Additionally, we may consider removing libraries which aren't maintained anymore (except they're complete, of course).