Structure save and load

• Ideally the folders are derived from a single point of truth, so when we make new experiments, we can just change one path.
• Theres paths specified in, e.g. model/trainer.py::save_model(), which required a bit of hacking around from me, and they didn't have the right relative path. Please standardize how the python programs are execute (e.g. python -m from project root) so these are consistent.
• [nice to have] The checkpoint / model save path could be exposed over argparse to a command line option, which makes training/loading a variety of models much easier.
• Review the first draft of the README.md
• Add missing and edit already existing sections