follow-ups after 761

[jeromedockes]

a couple of discussions that started in @glemaitre 's #761 but are not directly related and thus should not delay it:

• having the remainder columns in transformer_ always be column names, not indices comment
• or even not having a remainder anymore if we handle bools together with numerical columns comment
• always exposing the dtypes_ attribute, even when we don't autocast and thus dtypes_ is identical to the input's .dtypes comment

[Vincent-Maladiere]

having the remainder columns in transformer_ always be column names, not indices #761 (comment)

+1 and this is happening with scikit-learn/scikit-learn#27533

or even not having a remainder anymore if we handle bools together with numerical columns #761 (comment)

+1 and I'm wondering if they are edge cases / rare types that we should also handle.

always exposing the dtypes_ attribute, even when we don't autocast and thus dtypes_ is identical to the input's .dtypes

+1 on this one for sure.