ENH: Allow category-wise Fleiss' kappas.

[jseabold]

The API is a bit of a mess with many conditional returns.

Probably worth adding a bunch return and maybe a separate function with bells and whistles and deprecating the old one maybe?

Includes #9241, so it should go in after that. Will leave as draft until that's merged.

• tests added / passed.
• code/documentation is well formatted.
• properly formatted commit message. See
  NumPy's guide.

[josef-pkt]

Thanks @jseabold

We have Holder and HolderTuple which are now returned by most newer functions.
HolderTuple is backwards compatible with a tuple return (allows indexing and assigning to tuple) but not when previous return was only one value.

For transition and ambiguous cases where extra information might not be a priority, I had used return_results=False as keyword.
Otherwise, I now often return always extra results (if they are cheap) and intermediate results in a Holder instance.
If extra information is more expensive, I have been adding them recently as methods to a subclass of Holder, similar to the model results instance.

The Holder classes have a __repr__ that list all attributes with values.

Aside: Holder and HolderTuple do not subclass python dict or tuple, because I only wanted to support/maintain explicitly defined methods and not to have to worry about inherited methods from dict or tuple.

[jseabold]

Rebased on PR with results class returned.

Made it an error to request Randolph's Kappa with details. Not sure it makes much sense to assume a marginal 50/50 chance between a single category vs. all others?

If that makes sense, will need to remove that code path from the detail helper function, and I don't think I have a test yet for method = 'fleiss' and return_results = False.