Note: Built from PR Fix sparse y test #1248, should be considered after that.
Note2: Rebasing added some other PR's into this -_-

This PR fixes some typing issues that were present, notably the self.automl_ attribute is now created during __init__() rather than on the first call to fit, removing the _build_automl() function.

This was primarily to get rid of the typing warning that self.automl_.predict: None has no attribute predict(), as mypy can not establish that self.automl_ had been set in any methods.

This PR also does some other small typing cleanups.

• Union[SUPPORTED_TARGET_TYPES, spmatrix] -> SUPPORTED_TARGET_TYPES
• Move self.per_run_time_limit to be set at construction
• Explicitly marks AutoSklearnEstimator as an ABC
  • Was always functioning this way but was not marked as such
• AutoSklearnEstimator now relies on subclasses having two @abstractmethod so that validation only has to occur once and not in each subclass.
  • _supported_target_types
  • _automl_cls
• Removed method predict_proba from AutoSklearnEstimator, this should only be available in the subclass AutoSklearnClassifier
• Removed AutoSklearnRegressor.fit so it just uses the base class fit method. It was doing nothing but forwarding arguments.
• AutoSklearn2Classifer now explicitly sets attributes on the already constructed self.automl_, where previously it relied on it being constructed during fit

Main Change

# Now
class AutoSklearnEstimator(...):

    def __init__(...):
        self.automl_ = self._automl_cls(...) # Has type AutoML 
    
    def fit(...):
        ...

# Before
class AutoSklearnEstimator(...):

    def __init__(...):
         self.automl_ = None # Has type None
    
    def fit(...):
        if not self.automl_:
            self._build_automl()
        ...
        
    def _build_automl(self):
        self.automl_ = self._get_automl_cls()(...)

Extra Context

In general, as estimators are build in a two stage process, during __init__ and then remaining variables in fit, we end up with many attributes set to None during the __init__ phase. This causes problems with mypy's typing as it can not garuntee that it will not be None when the are used later.

There are two solutions to this I can think of:

• Move any attributes we can create to __init__ from fit. In general this is probably better practice and safer.

• Move attributes to a @property

  @property
  def attribute_wanted(self) -> T:
      if not self._attribute_wanted:
          self._attribute_wanted = ...
      return self._attribute_wanted

  This second solution works fine when we can use the combo @property attributed_wanted with self._attributed_wanted.
  It has the issue when we want @property _attribute_wanted and self.__attributed_wanted.

  While we could have the reference as __attributed_wanted, with the double leading underscore, this invokes pythons name scrambling and subclass will not be able to access this attribute directly and will be require to use the @property _attributed_wanted. This may or may not be a problem.