[lightning] Add support to continue training

[karl-richter]

Note: This feature depends on the introduction of Lightning as model wrapper.

Desired functionality
The training of a model can be continued. The fit function can be called a second (arbitrary number) of time(s) to tune the model performance.

Current state

• Test test_continue_training added under test_utils.py
• Parameter continue_training is supported by fit(...)

Required changes
Currently the learning rate scheduler throws a ValueError, indicating that the number of steps is above the originally defined number of steps (which makes sense). The new max_step number likely needs to be adjusted somewhere.
Further, the learning rate scheduler of the model likely needs to be switched when continuing training.

[karl-richter]

In progress, co-development with @Schludel.

[Schludel]

TODO: Update total steps of OneCycleLR with new total steps in if-statement

[karl-richter]

To decide: Should we continue using the OneCycleLR in case we continue training the network or should we rather use a simple decaying lr scheduler instead that uses the last learning rate from the previous lr scheduler?
Todo: Run experiments to explore