add resource sharing for efficiency

[sfluegel05]

As of now, all models work on their own - each is called separately and initialised separately. This leads to some inefficiencies. For instance, ChemLog extra and the smoothing both load ChEBI from a chebi.obo file and construct a hierarchy graph. They don't even use the same chebi.obo file.

In general, it is not possible to share resources between arbitrary models, but in some cases, this might be useful.

Todo

• Provide some standard resources by the ensemble that might be needed for different models (e.g., a chebi.obo file, a hierarchy graph, rdkit mol objects)

[sfluegel05]

As of now, the ensemble loads a chebi graph (i.e., a networkx graph with all ChEBI classes and is-a relations between classes (including transitive ones) and distributes it to all models.

Another useful thing might be to move SMILES preprocessing to the ensemble. This would also help with identifying faulty SMILES (currently, each model notices, independently, that the SMILES is faulty, and the ensemble has to figure out why it got no predictions - it would be easier if the ensemble could check a SMILES string's validity before passing it on).