Video generation fix, Checkpoint restore on training and some code cleanup.

[jahnog]

• Fixed ffmpeg video generation. With moviepy 1.0.1 from the conda-forge channel, specifying a video file to import it's audio, also imported the original video track and the process failed in the middle with an exception.
• Added a parameter to style.py so the training can resume from the last saved checkpoint and continue optimizing from there.
• Made some code cleanups to be able to run the scripts in Python 3.7.7