Tune atom and bonds featurisers to achieve best results for predicting AdrA1A_PCHEMBL_VALUE

[mkviatkovskii]

Is blocked by #602 and #622.

ToDo

1. Take minimal set of featurisers (atom number and bond order) as a basis.
2. Write a simple engine that allows to add one or more featurisers to the mol2graph and concatenate them.

Next step is pretty creative, there could be different approaches. Just one of them as an example:

1. Write test code that automatically iterates over list of featurisers and adds them one by one.
2. Take minimal mol2graph and iterate over all featurisers to test what pair of featurisers act better: atom number and valence, atom number and mass etc.
3. Choose best model for two featurisers, then go back to step 3 and add one more featuriser from the rest of possible etc.
4. Stop when model is not achieving better results anymore.

After that add a ready preset of featurisers (as it's done in DGL-LifeSci).