Implement train_step instead of overwriting fit

[tabergma]

Proposed changes:

• related to https://github.com/RasaHQ/research/issues/97
• (Training is currently not working due to 'fit' takes a dataset with increasing batch size tensorflow/tensorflow#41019 - we have an open PR for tensorflow. We probably copy over some code in the meantime.)

Status (please check what you already did):

• added some tests for the functionality
• updated the documentation
• updated the changelog (please check changelog for instructions)
• reformat files using black (please check Readme for instructions)

[tabergma]

The results of the model regression tests do not match the results I got when I'm testing locally. Need to investigate that further.

[joejuzl]

I profiled the memory usage while running tests/core/test_policies.py on master, and on this branch.

master:

train_step (commit 4ea8d8 - current head )

train_step (commit 07d63c - before keras clear_session)

Observations:

• Takes much longer to run on train_step.
• Way higher memory usage on train_step, even at the same relative point in time.
• The Keras clear_session seemed to help somewhat.
• All show a continuous increase in memory usage, suggesting a memory leak.

[Ghostvv]

@joejuzl could you please give it a final review?

[joejuzl]

#8085 should help to reduce the memory usage of the policy tests and stop the CI tests failing.

[Ghostvv]

@joejuzl thank you for updating the tests. Could you please merge your changes into this branch? There are some merge conflicts