Early stopping based on unique generated SMILES

[NicklasOsterbacka]

This PR adds a new stage parameter (max_smiles) that stops generation once the number of canonicalized generated SMILES reaches max_smiles. This also changes the behaviour of learning.py to add only valid SMILES to smiles_memory.

This allows for the running of the Practical Molecular Optimization benchmark through REINVENT4 by setting max_smiles to 10000.

[NicklasOsterbacka]

I have additionally added a new terminator that stops execution after the top-k scores no longer improve after N epochs. This is primarily meant for the Practical Molecular Optimization benchmark in conjunction with unique SMILES-based early stopping, but might be interesting as a general alternative to checking mean scores as well.

I refactored the convergence check in learning.py to accommodate this change; it should in principle be a bit easier to extend it with more terminators, if need be. I also added two new stage parameters, topk and patience, as settings for the new terminator.