Create CatMatrix from codes and categories

[MarcAntoineSchmidtQC]

See this issue on Glum to understand the reasoning.

Checklist

• Added a CHANGELOG.rst entry

[MarcAntoineSchmidtQC]

@lbittarello, if you have time I would love to get your feedback. It removes the _Categorical class that you added during the polars PR.

[stanmart]

Looks good. I feel that a separate constructor, e.g. CategoricalMatrix.from_codes (like pandas.Categorical.from_codes) might be a bit clearer than the meaning of cat_vec being dependent on another argument, but I'm on board with this solution, too.

[lbittarello]

I'm happy with the option to create a categorical matrix from codes. Indeed, it was my preferred option. :)

However, this PR has an unintended consequence. If we use a Polars array to instantiate a categorical matrix, the cat property returns a Polars array. If we then subset this categorical matrix, the subset will no longer remember that the original input was a Polars array, because the _input_dtype attribute will be set to the dtype of the array of categorical codes. At that point, cat will return a Pandas series.

The _Categorical container ensures that, no matter how many times we subset a categorical matrix, we'll always remember the input type and cat will always return an appropriate series.

I'm not sure that we care about it. I added a deprecation warning to cat to that we'd be able to take categorical codes in the future without these complications.

Alternatively, we could add an input_dtype instantiation parameter for __getitem__ to set, but it seemed a strange attribute for users to have access to.

[MarcAntoineSchmidtQC]

If we deprecate the cat property, I don't think we should be thinking about expanding its scope. Polars support is not released yet, so in the stable version cat is a pandas.Categorical. I would be happy to simply return a pandas.Categorical when possible and otherwise raise an error saying that this method is not supported with a non-pandas backend. This is fully backward compatible.

[MarcAntoineSchmidtQC]

Of course, I can add a more obvious error message instead of relying on python spitting out an error because it doesn't know what pd is.

[stanmart]

Would adding an _input_dtype argument to CategoricalMatrix's constructor solve this? __getitem__ sees self and can pass this property to the newly created matrix.

[lbittarello]

I would be happy to simply return a pandas.Categorical when possible and otherwise raise an error saying that this method is not supported with a non-pandas backend.

That's fine by me. :)