Expand experiments

[jakub-piw]

We need to expand experiment framework with additional sttaistics like:

• number of models in hellsembles
• coverage of datasets by component models
• progressive performance of hellsemble on whole dataset (with 1, 2, 3,... models in ensemble) to check if the score is improving correctly
• performance of routing model
• performance of components models on corresponding subsets
• other ideas