inconsistent treatment of None and np.NaN in SimpleImputer

Doing constant imputation treats only the "missing_value" as missing, so a None by default stays there:

from sklearn.impute import SimpleImputer
import numpy as np
X = np.array([1, 2, np.NaN, None]).reshape(-1, 1)
SimpleImputer(strategy='constant', fill_value="asdf").fit_transform()

array([[1],
       [2],
       ['asdf'],
       [None]], dtype=object)

However, using strategy='mean' coerces the None to NaN and so both are replaced:

SimpleImputer(strategy='mean').fit_transform(X)

array([[1. ],
       [2. ],
       [1.5],
       [1.5]])

I don't think the definition of what's missing should depend on the strategy. @thomasjpfan argues that the current constant behavior is inconvenient because it means you have to impute both values separately if you want to one-hot-encode.

It seems more safe to treat them differently but I'm not sure there's a use-case for that. This came up in #17317.
I think this only matters in these two, as other imputers don't allow dtype object arrays.

Start at the SimpleImputer implementation and its existing tests for object arrays, then reproduce the examples with strategy='constant' and strategy='mean'. Review the discussion around #17317 and determine the intended treatment of None and np.NaN; done means the behavior is consistent with that decision and covered by regression tests.