Update GLN environment to Python 3.9

[kmaziarz]

Unlike all other supported single-step models, GLN uses a separate environment defined with docker, as it depends on building rdkit from source. GLN is also the only model using Python 3.7 - a version which reached its end-of-life last year. This PR keeps the docker-based setup, but updates the GLN code and environment to Python 3.9, so that in a future version of syntheseus we can raise the required Python version to 3.9 and get rid of backward compatibility hacks.

[AustinT]

Seems good to me- I didn't try running it myself since I don't actually have docker installed, but the changes on your custom GLN repo look pretty minimal (just changing a single pytorch function call). If the tests pass then I think everything should be ok?