Add prediction quality monitoring

[h-huss]

An automatic assesment of the prediction quality could be a very helpful tool for further tuning. Ideally, predictions should be run for a number of repositories on each commit, similar to contineous integration. This could help to state whether a particular change is useful.

To get meaningful results, these predictions should be done on a large number of repositories with enough computing power. It might be hard to achieve this on Travis.

Furthermore, machine learning is a stochastic process - random fluctuations in result quality might overshadow the effects of changes.

The actual evaluation implementation was started in 4ac580e.