Refactor Task into single "training loop"

[vitusbenson]

Originally the library was intended to be able to handle multiple tasks. In the end it seems all models can run with the spatio-temporal task. Hence it is unnecessary to keep this abstraction.

Goal:

• Simplify code, remove stuff that is related to the possibility of multiple tasks

Risks:

• Backward compatibility