Type estimators

[eddiebergman]

This is a pretty big PR aimed at doing a simple thing, remove estimators.py and automl.py from the mypy ignore list. In Progress, notes on changes are TODO. I'll resolve conflicts once v0.15 is out. I can also split this into multiple smaller PRS to make it easier if needed.

Tests still need to be updated to accomodate changes.

There were 168 typing errors :) Some of them were actual possible bugs based on order of things being called and parameters set.

Major points:

• Made AutoML an ABC. This removes some failure cases such as when task type is not defined or is_classification is misspecified, see Tests for the AutoML class relying on is_classification=false even when it is a classificaiton task, crash when corrected #1212. AutoML now relies on its subclasses specfying things which was also simplified greatly with just class variables.
• Also removed the methods which just forward all the args with is_classification=True/False since those parameters were removed as well.

class AutoMLRegressor(AutoML, RegressorMixin):
    _task_mappings = {...}
    is_classification = False
    
class AutoMLClassifier(AutoML, ClassifierMixin):
    def predict(...): ...
    def predict_proba(...): ...

---

Made the AutoSklearnEstimator smarter with respect to types in a similar fashion, notably it's smarter around what it retunrns through the use of a Generic in the main class and providing those types in the subclass. This mainly means that code editors will know if predict_proba will be available or not and that fit will return the the right estimator and not just the abtract AutoSklearnEstimator.

Self = TypeVar("Self", bound="AutoSklearnEstimator")

TParetoModel = TypeVar("TParetoModel", VotingClassifier, VotingRegressor)
TAutoML = TypeVar("TAutoML", bound=AutoML)

class AutoSklearnEstimator(ABC, BaseEstimator, Generic[TAutoML, TParetoModel]): ...

    # Knows it returns the same type as self, AsklearnClassifier or AsklearnRegressor
    def fit(self: Self, ...) -> Self: ...
    
    # Knows if its a AutoMLClassifier, AutoMLRegressor
    def automl() -> TAutoML: ...
    
    # Knows that the pareto models are a VotingClassifier/Regressor 
    def get_pareto_set() -> Sequence[TParetoModel]: ...

These are then specified in the subclass as

class AutoSklearnClassifier(AutosklearnEstimator[AutoMLClassifier, VotingClassifier]): ... 
class AutoSklearnRegressor(AutosklearnEstimator[AutoMLRegressor, VotingRegressor]): ... 

---

• Many things that are set in fit are now wrapped in a property, ie. self._logger or self._task and raising a NotFittedError as sklearn would. This is because using them in other methods would correctly warn something like "self._task could be "None" if trying to call methods relying on fit to have been called first.

class AutoMl:

    @property
    def task(self) -> int:
        if self._task is None:
            raise NotFittedError("`task` has not been set, please call `fit` first")

        return self._task
        
    @property
    def input_validator(self) -> InputValidator:
        if self._input_validator is None:
            raise NotFittedError(
                "`input_validator` has not been set, please call `fit` first"
            )

        return self._input_validator

---

• Make transform, get_cost_of_crash smarter with @overload, it now knows it's return type correctly based on type of input.

@overload   # Not the type for y is None, indicating None returned
def transform(self, X: XType, y: None = None,) -> tuple[XType, None]: ...

@overload  # And here, it's different
def transform(self, X: XType, y: YType) -> tuple[XType, YType]: ...

def transform(self, X: XType, y: YType | None = None) -> tuple[XType, YType | None]:
    ...

---

• Run pyupgrade on a few files I touched

• Simplified the datacompression things into a class, the typing caught some weirdness when datacompression was on but memory_limit wasn't set.