[ENH] make TabularToSeriesAdaptor compatible with sklearn transformers that accept only y, e.g., LabelEncoder

[fkiraly]

This PR makes TabularToSeriesAdaptor compatible with sklearn transformers that accept only y, e.g., LabelEncoder.

If provided with such a transformer, it plugs in the outer X to the inner y, along the lines of the a suggested design resolution at the root of the problem in sklearn, discussed in #5867.

Towards #5867

FYI @yarnabrina

Question: do you know of a transformer that has (a) only y, and (b) accepts float input? This is to ensure we can test it with the default framework, which currently does not make distinctions in input types.

[yarnabrina]

FYI @yarnabrina

Question: do you know of a transformer that has (a) only y, and (b) accepts float input? This is to ensure we can test it with the default framework, which currently does not make distinctions in input types.

Can you not use LabelEncoder itself, with a categorical variable whose levels are numbers?

I don't know of other transformers specifically for real variables.

[fkiraly]

Can you not use LabelEncoder itself, with a categorical variable whose levels are numbers?

That would result in failure (I guess but have not tried), as currently all transformers are tested with all data scenarios, and these are float.

We could do sth similar as #5964 for transformations, but then we'd have to introduce tags for the types both in transformers and scenarios, and ensure there is logic to match them (in is_applicable).

[fkiraly]

Hm, I think this cannot be properly tested without having some version of #5886 functioning first. At least, that is required to add #5867 as test case.