Refactor argparse for python kwargs

[vitusbenson]

Currently in many parts argparse is used to give hyperparameters and options to modules. This is mainly because in its early time, PyTorch Lightning was built this way. A more pythonic way is to just use normal keyword arguments for functions. These can also be nicely passed by a dict func(*args, **kwargs). The benefit would be that it is easier to use the library from inside Jupyter notebooks.

Goal:

• Replace argparse hyperparameters by python kwargs

Culprits:

• Probably can not really do this for trainer arguments (?)

Risks:

• Backward compatibility