Extending Auto-sklearn can be incompatible with parallel execution

[mfeurer]

When running Auto-sklearn extended with a classifier/regressor etc with n_jobs>1 it can happen that the classifier/regressor Auto-sklearn was extended with is not available in the subprocess. Therefore, all function evaluations will crash.

This can be circumvented by calling add_classifier etc in the subprocesses executing the machine learning runs. Sometimes, this is done automatically when for example the call to add_classifier is done on the script's highest level and is not guarded by __name__ == "__main__". However, it would not work in a notebook for example.

Potential solution:

1. Move the _addons attribute every component type has to a global addons registry.
2. Pass all registered addons to the workers
3. When doing a function evaluations, check whether addons need to be registered on the worker and do so

[mfeurer]

This is fixed in #1290. I'm keeping it open because we need to document that this still will not work in a notebook because the component will not be importable in a worker.