RFC _convert_container

[Charlie-XIAO]

Follow up of #28521, see #28521 (comment).

This PR splits constructor_name into constructor_type and constructor_lib, and adds two parameters sparse_container and sparse_format to reflect dense versus sparse. /cc @glemaitre

Mapping from the original constructor_name to the specifications in this PR

	constructor_type	constructor_lib	sparse_container	sparse_format
Default		"pandas"	None	"csr"
"list"	"list"
"tuple"	"tuple"
"slice"	"slice"
"array"	"array"
"sparse"/"sparse_csr"	"array"		"matrix"
"sparse_csc"	"array"		"matrix"	"csc"
"sparse_csr_array"	"array"		"array"
"sparse_csc_array"	"array"		"array"	"csc"
"dataframe"/"pandas"	"dataframe"
"polars"	"dataframe"	"polars"
"pyarrow"	"dataframe"	"pyarrow"
"series"	"series"
"polars_series"	"series"	"polars"
"index"	"index"

Some rules of when each parameter is applicable

• sparse_container has no effect unless constructor_type="array"
• sparse_format has no effect unless constructor_type="array" and sparse_container is not None
• constructor_lib has no effect unless constructor_type is one of "dataframe", "series", and "index"
• min_version has no effect unless constructor_lib is used
• column_names (renamed from columns_name) has no effect unless constructor_type="dataframe"
• categorical_feature_names has no effect unless constructor_type="dataframe"

Constraints on the shape of the input container

• constructor_type="slice" requires 1-dimensional container with exactly 2 elements
• constructor_type="dataframe" requires 2-dimensional container (previously it converts (n,) shape to (n, 1) shape)
• constructor_type="series" requires 1-dimensional container (previously it creates series with each element being a list when given 2-dimensional container)
• constructor_type="index" requires 1-dimensional container (similar to "series")

[github-actions]

✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: 5d89d71. Link to the linter CI: here}

[jeremiedbb]

Thanks for the PR @Charlie-XIAO. Here are some comments. I'm +1 for the refactoring of _convert_container.

[glemaitre]

Once the CIs are green, I'll review ;) (I am buying a bit of time).
But indeed, I really like the refactoring. Thanks for taking care of it @Charlie-XIAO.

[Charlie-XIAO]

Now I have resolved all suggestions from Jérémie, and CI is green :)

[glemaitre]

I'm going to resolve the conflicts and make a review.

[glemaitre]

A couple of comments before to look at the tests.

[glemaitre]

If I get it right, we are going to convert any dataframe to an array at this stage. I'm not sure this is wise since I could imagine converting a dataframe pandas into a dataframe polars but I would not need to go back to numpy, isn't it?

[glemaitre]

In some way, I would as much as possible delay calling asarray to be sure that we actually need it.

[Charlie-XIAO]

I didn't think that much of performance given this is a testing utility, but your point is valid. This would mean refactoring and I will do it when I get some full block of time :)

[glemaitre]

I didn't think that much of performance given this is a testing utility, but your point is valid.

I'm a bit more scared about the dtype casting where will enforce a single dtype while we might be interested in dataframe with heterogeneous dtype and just trying to test for pandas and polars. Passing by NumPy in the middle could make it a mess :)

[Charlie-XIAO]

Very true, missed this point.