Consider updating smiles.py to avoid discrepancy between RDKit hybrid…

[bmp192529]

…izations and x_map

I encountered an error in training my own model where some SMILEs strings contained atoms classified as SP2D by RDKit. In the end, it turned out there were only 3 such SMILEs strings in my data that were preventing interpretation of thousands of others. This works at least as a temporary solution.