Support custom BaseSearchCV subclasses by accepting param_grid or param_distributions

[axif0]

Removes the hard-coded check for only GridSearchCV and RandomizedSearchCV.
Now, any BaseSearchCV subclass is supported as long as it provides either param_distributions or param_grid.
The n_jobs safety check remains unchanged.

I tried this code, to reproduce the warning -

class CustomSuccessiveHalving(sklearn.model_selection._search.BaseSearchCV):
    def __init__(self, estimator, param_distributions):
        self.param_distributions = param_distributions
        self.estimator = estimator
        super().__init__(
            estimator=estimator,
            scoring=None,
            n_jobs=None,
            refit=True,
            cv=3,
            verbose=0,
            pre_dispatch='2*n_jobs',
            error_score='raise',
            return_train_score=False,
        )
    
    def _run_search(self, evaluate_candidates):
        from sklearn.model_selection import ParameterGrid
        param_grid = ParameterGrid(self.param_distributions)
        evaluate_candidates(param_grid)

if __name__ == "__main__":
    extension = SklearnExtension()
    
    base_estimator = sklearn.ensemble.RandomForestClassifier()
    param_distributions = {
        "max_depth": [3, 5],
        "n_estimators": [10, 20]
    }
    
    custom_hpo = CustomSuccessiveHalving(
        estimator=base_estimator,
        param_distributions=param_distributions
    )
    
    extension._prevent_optimize_n_jobs(custom_hpo)

Closes: #6