Add a toy problem

[vitskvara]

Add a toy dataset on which we can benchmark methods. Ideally, a working detector should have (almost) perfect prediction accuracy on it.

[masenka31]

I really like the toy problem from Model-Based Learning for Point Pattern Data (Figure 7) to test for cardinality influence on detection: http://prntscr.com/10zwlwf
Neural Statistician was able to score AUC=1 for all 3 cases with some score functions.

The toy problem from One-Class Support Measure Machines for Group Anomaly Detection (or a modification of such toy problem) looks good for "semantic" anomaly detection: http://prntscr.com/10zwmd2. It can be also modified to have different cardinalities as well.

[vitskvara]

Nice, I like both problems.