RFC/MAINT: refactor and common tests for estimator supporting string (object dtype)

[glemaitre]

Following #13046, we have redundant code in estimator which should support string dtype (object dtype). Those estimators should make similar validation and raise consistent errors and have (a) common test(s).

The estimators which are involved will be:

• SimpleImputer
• MissingIndicator
• OneHotEncoder
• OrdinalEncoder

I open this issue such that we can agree on the way to refactor this part.

[amueller]

estimator tags should allow us to specify different training data for the estimators. I think that's the way to go.

[glemaitre]

So you propose to make a common test based on the X_types.

@amueller Do you need more review on #8022, it seems almost good to be merged last time I checked?

[amueller]

No, I propose to make all common tests use the X_types to figure out what training data to use. Right now most checks can't be performed for, say, CountVectorizer (or OneHotEncoder) because we can't train it because all the tests use float data for fitting.

re review: Not sure, I think @rth was reviewing.

[jnothman]

CountVectorizer, DictVectorizer are also special because their input is 1d