A training process can be started using a worker version in a created or error state

You can reproduce this from the frontend: first follow the steps of frontend#1181 (closed), then finish filling out the other form fields and just click on the start button. With some worker versions that have a Docker image set (not worker versions that came from DataSources for example), the process will be started just fine, and then splendidly crash with a System error from Ponos because the Docker image does not actually exist. Crash example in preprod

An extra ValidationError is missing in arkindex.training.serializers.StartTrainingSerializer…