CLN: remove fastpath & verify_integrity from constructors

[jreback]

this is a top-level issue to remove the fastpath and/or verify_integrity keyword args from constructors from public interfaces, classes such as Categorical, Series, andMultiIndex. If these are removed all once from a particular class, e.g. Index/MultiIndex for verify_integrity, but are still needed for an internal API, then can add an internal constructor _from_fastpath which can serve instead (e.g. see pandas.core.dtypes.dtype.CategoricalDtype for an example of this) There may be some re-factoring required to keep things DRY.

Most of these will actually need to be deprecated.

• Index (fastpath)
• Index (verify_integrity in subclasses)
• MultiIndex (verify_integrity)
• pd.concat/.append (verify_integrity)
• Series (fastpath) -> DEPR: Series constructor fastpath keyword #55466
• Categorical (fastpath) -> DEPR: Categorical fastpath #52472

[jreback]

xref #20017, cc @PoppyBagel

[TomAugspurger]

Pushing to the next release if that's OK.

[jorisvandenbossche]

I added it to 0.24.0 instead of the big pile of "Next major release"

[spacesphere]

@jreback, sorry for the delay with it. Could you, please, clarify some points.
Does this mean that we need to deprecate these arguments first?
Also, unlike with fastpath, it doesn't seem totally clear to me how to deal with verify_integrity (especially in the case of pd.concat/.append). And should it be treated as fastpath in sense of naming?

[TomAugspurger]

Does this mean that we need to deprecate these arguments first?

Yep.

I'm not sure about deprecating verify_integrity.

[jorisvandenbossche]

I think verify_integrity is a potentially useful public argument, so also not sure about deprecating it.
In any case, I think fastpath is more important, as that is really exposing internal details.

[jbrockmendel]

If we deprecate fastpath in the Series constructor, for perf we'd want to have those cases go through Series._from_mgr, or more specifically self._constructor_sliced._from_mgr. But this will break whenever _constructor_sliced doesn't return a class (eg geopandas IIRC).

There are a couple of potential workarounds

1. deprecate allowing non-class _constructor_sliced. deprecation path might be tough. downstream could maybe keep old behavior by implementing __new__?
2. implement DataFrame._from_mgr_sliced that subclasses could override.