Reference Issues/PRs

Fixes #19071.
In SimpleImputer strategies mean/median you can use input arrays with dtype is object. Imagine numeric data in the input array, except for missing values. In object arrays, you have a lot of flexibility on how to encode missing values. You can use np.nan, None, a string constant or anything else. There are two scenarios which are described in the issue #19071:

1. There is more than one type of missing value in the input object array, one of them is specified as the missing_values parameter of the SimpleImputer. In this case the mean/median imputation should fail, because there are still non-numeric values in the input array other than missing_values which is masked. See SimpleImputer, missing_values and None #19071 for an example of not failing.
2. There is just one type of missing value in the input object array and that is specified as the missing_values parameter of the SimpleImputer. In this case the mean/median imputation must not fail and should calculate mean/median from the remaining values of the input array. See SimpleImputer, missing_values and None #19071 for an example of this failing with missing_values=None.

What does this implement/fix? Explain your changes.

In the _validate_input method, I keep the dtype as object.
In the _dense_fit method I try to convert the object array into float64 to prepare for mean/median calculation. If the not-masked values still contain some non-numeric values (scenario 1), ValueError is raised with the following informative error message:

Non-numeric values other than missing_values={missing_values}, showing {num_show}/{num_notmasked_nan}: {examples}

where the number of examples is limited to three.

Any other comments?

A comprehensive set of unit tests is added, which covers both scenarios.