Allow data preprocessing

[hleepfl]

Mother issue about offering data pre-processing inside the platform.
For now, connecting data and clicking the "train" button is very easy. On the other hand, the data that gets connected has to be at least partly pre-processed already.
To harmonize the DISCO experience, we want the entire process from Step 1 (having raw data and wanting to coordinate with other users to train a model) to Step n (inference on new raw data) to have a consistent requirement in terms of technical knowledge (close to zero).
One of the things that is missing to achieve this is transforming raw data to pre-processed, ml-ready data.

Steps

• Design (do we want pre-processing pipelines to be part of a task's definition ? should they be defined separately then assembled with a task later ? and more)
• Create the related issues and link them to this one
• Implementation

Related issues

#649, #1062, #1063

[JulienVig]

My impression is that the task creator would define the preprocessing pipeline, which would then be applied automatically for each participant. Disco could implement a number of preprocessing functions that can be used as building blocks in the training task specification.