CLN: use _values_for_argsort for join_non_unique, join_monotonic

[jbrockmendel]

With the .copy() removed from Categorical._values_for_argsort, ea_backed_index._data._values_for_argsort() matches ea_backed_index._ndarray_values in all extant cases.

cc @jorisvandenbossche @TomAugspurger need to confirm

a) this is an intended-adjacent use of _values_for_argsort, and not just a coincidence that it matches extant behavior
b) the .copy() this removes from Categorical._values_for_argsort is not important for some un-tested reason

xref #32452, #32426

[TomAugspurger]

I believe the only requirement on values_for_argsort is that it's a monotonic transformation (it preserves the ordering).

I'm not sure why that copy was there.

[jbrockmendel]

I believe the only requirement on values_for_argsort is that it's a monotonic transformation (it preserves the ordering).

It isn't clear to me what distinguishes this from values_for_factorize, as the docstring there also says An array suitable for factorization. This should maintain order and be a supported dtype. (the "supported dtype" bit isnt in the values_for_argsort docstring, but seems implied since it returns an ndarray)

This also came up in #30673 (attempt to implement value_counts in terms of EA methods). See also #32412.

Do we know of any 3rd-party EAs where _ndarray_values, values_for_argsort() and _values_for_factorize()[0] are meaningfully distinct?

[jreback]

are there any user facing things that this now allows? e.g. joins on EA?

[jbrockmendel]

are there any user facing things that this now allows? e.g. joins on EA?

behavior is unchanged

[jreback]

thanks, pls followon with consoilidations when you can

[jorisvandenbossche]

My feeling says that this should use _values_for_factorize, as joining is a factorize-based algo?

[jbrockmendel]

My feeling says that this should use _values_for_factorize, as joining is a factorize-based algo?

I think you're right, but ATM _ndarray_values matches values_for_argsort for all our Index-backing EAs, but values_for_factorize()[0] is slightly different for DTA/TDA (_data vs asi8). If we determine that we can change DTA/TDA _values_for_factorize (possibly as part of the discussion in #32586) then ill switch over these usages.