Handling of ExecutionError in generators

Current situation

A generator job may publish an ExecutionError event while handling a GeneratorCommand message. Historically, this was then cancelling the workflow (the generator we used most was not specifying complete metadata, hence that was the best thing we could do).

The generator has been fixed. So now we should handle those failures nicely: it should fail the job, and then possibly the parent job, and so on, depending on continue-on-error and such.

If a generator wants to cancel a workflow, it can publish a WorkflowCanceled event. It's a user process, it should rely on user APIs.

Desired outcome

Remove the previous now incorrect behavior, implement a proper behavior.