Sparse matrices, code reorganization, pip installable project, using pickled datasets

[lambdaofgod]

I've added following changes:

• use sparse matrices. This allows for loading bigger graphs, way bigger than it was possible using dense matrices
• setup.py was added and lorentz module was extracted and moved to lorentz_embeddings/lorentz so that project is pip-installable and Lorentz module can be used outside
• added load_dataset function to datasets so that user can specify sparse matrix pickle filename as dataset argument
• code was reorganized to run experiment in main.py

[theSage21]

@lambdaofgod thanks for the awesome work! I currently don't have time to test this PR so could you please also update the examples using the code changes in this PR? A new example showcasing the change in magnitude available due to sparse matrices would be wonderful alongside regenerated examples using the new code.

I think once the examples have been updated we can merge this since they would demonstrate that the code works as expected.