Remove legacy training code

[thodkatz]

As with the newest attempt of adding neural network training in tiktorch #225, we should remove any legacy old-designs of training functionality.

The #225 has already taken care some parts of it. But still some leftover code is remaining.

For example:

• Add training service #225 (comment)
• Commands UpdateDatasetCmd and SetMaxNumIterations
• the models dir
• the scripts dir
• the configkeys
• the dummy server contains a training interface
• the tiktorch types
• the session types
• the tiktypes

As I was trying to design a new training approach, I have tried to be inspired by the existing leftover parts, but at the same time I was struggling to follow it, as it seems a lot of assumptions were made from the bioimageio side as well. Thus, some of them are remained, but probably we should remove them, and try to follow one consistent approach.