Make the TrainingInformation polymorphic

[Grim-bot]

Is your feature request related to a problem? Please describe. Right now, the TrainingInformation is too flat: it contains many optional arguments that shouldn't be optional for certain schemes, and shouldn't be possible for others. For instance:

• inputColumns and outputColumns are necessary for tabular data and have no effect for image data. IMAGE_H and IMAGE_W are the opposite.
• minimumReadyPeers is only needed for decentralized

Describe the solution you'd like I would like to see a trainingInformant with more depth, where it's impossible to specify inputColumns for image data and where it's impossible to not specify it for tabular data.

Describe alternatives you've considered We could think about using classes or more complicated interfaces...

Additional context This also affects what we can do with the UI