Add support for pandas 1.4.0

Looks like the root cause may be that converting a nullable type to Categorical preserves the nullable types in the pandas Categorical categories. On the top is the latest pandas and on the bottom is 1.3.4 pandas.

This causes np.asarray calls in our imputers to introduce the new pandas null value place-holder, <NA> which throws off sklearn:

_Originally posted by @freddyaboulton in https://github.com/alteryx/evalml/issues/3272#issuecomment-1020261986_

This change in behavior breaks tests in three places:

• In our simple imputer, we convert NaturalLanguage to Categorical in order to support imputing natural language via the most_frequent strategy. (test_simple_imputer_supports_natural_language_constant)
• In our imputer tests, we test running the imputer on a dataframe that had a nullable int column converted to categorical (test_imputer_woodwork_custom_overrides_returned_by_components)
• In our tests for EmailFeaturizer, and URLFeaturizer, this causes the categorical features created from Email and URL logical types to have nullable types in the categories because the physical type for Email and URL is text. See (test_ft_transform_primitive_components.py::test_component_fit_transform[component1-make_data_email_fit_transform_missing_values-make_answer_email_fit_transform_missing_values-make_expected_logical_types_email_fit_transform_missing_values])

Additionally:

• There is a warning being emitted by xgboost using pd.Int64Index in data.py (https://github.com/dmlc/xgboost/blob/v1.5.0/python-package/xgboost/data.py#L250). This causes test_xgboost_catch_warnings_label_encoder to fail.

[chukarsten]

Looks like the XGBoost warnings are remedied, per https://github.com/dmlc/xgboost/pull/7595! Just need to wait for a release from XGBoost.

[thehomebrewnerd]

@chukarsten Just a heads up, Featuretools doesn’t yet fully support pandas==1.4.0 either. It might work for you though, but no guarantees at this point.

https://github.com/alteryx/featuretools/issues/1865