Faster and more diverse de novo molecular optimization with double-loop reinforcement learning using augmented SMILES