Now that we propagate the enabling/disabling to descendants, we no longer have to check the disabled_observer Set on each superclass of the model class. This was causing a bug when disabling all observers at a superclass level and then enabling an individual observer at a subclass level.
Plus the logic is simpler now :).
This reverts commit ad62f19287 because the current
subclasses implementation leaks memory in development.
Instead of keeping an array of subclasses, the better solution is to
ask the parent if an observer is being disabled or not.
