MAINT: Remove unsafe unions and ABCs from return-annotations

[BvB93]

Backport of #18885.

Per the title, this PR removes unsafe unions and abstract baseclasses from the return annotations,
e.g. functions that currently have one of the following patterns.

from __future__ import annotations
from typing import Any, Sequence, Any as A, Any as B

def func1(*args: Any, **kwargs: Any) -> A | B:
    pass

def func2(*args: Any, **kwargs: Any) -> Sequence[Any]:
    pass

The Problem

The problem with returning a Union (or, almost equivalently, an abstract-ish baseclass such as generic)
is that any and all operations performed on a union must be compatible with all of its members.
For example, operations that are exclusive to either np.float64 and np.ndarray are thus not allowed
to be executed by np.float64 | np.ndarray, unless the union is narrowed down via an explicit isinstance
check a priori.

While returning a union thus adds some a form of simplicity to the annotations (as we don't have to
distinguish between 0D and ND array-likes), the Union type is simply not suited for what we're trying
to describe here (xref python/mypy#1693). This would be a different story
for a hypothetical UnsafeUnion type, one where operations must be compatible with any member
of the union. Such type does not exist though.

from __future__ import annotations
from typing import Any
import numpy as np

array: np.ndarray[Any, np.dtype[np.float64]]
out = np.isneginf(array)

if TYPE_CHECKING:
    # note: Revealed type is 'Union[numpy.bool_, numpy.ndarray[Any, numpy.dtype[numpy.bool_]]]'
    reveal_type(out)

for i in out:
    # error: Item "bool_" of "Union[bool_, ndarray[Any, dtype[bool_]]]" has no attribute "__iter__" (not iterable)
    print(i)

# Only now will things
if isinstance(out, np.ndarray):
    for i in out:
        print(i)
else:
    print(out)

The Solution

The solutions implemented herein fall in either one of the following two categories:

• Simply set the return type to Any. This is a simple, but non-ideal solution, as it removes any type
  safety for objects returned by aforementioned functions. Nevertheless, there is a group of functions
  that still needs to-be updated for dtype-support anyway (e.g. those in np.core.fromnumeric),
  so setting their return to Any is by no means the worst thing that can happen.
  A second group of functions is those where the use of Any is simply a necasity, as the output type is.
  For example, determined by the value of string literals (see np.einsum). As a silver lining: the einsum
  problem in particular does seem like it can be resolved with relative ease via a future mypy plugin.
• Add additional an additional overload for 0D array-likes. This is the more thorough and permanent fix
  for the unsafe-union issue; it has been applied to the more recently annotated modules such as
  np.lib.ufunclike.
  There is however one important caveat here: as we currently lack shape-support (Typing support for shapes #16544) it is
  currently impossible to distinguish between 0D and ND ndarrays, and thus we are unable to describe
  the 0D-to-scalar casting that numpy aggressively performs on 0D arrays. While this should change
  once PEP 646 is live, in the mean time users will have to settle for a typing.cast call or a # type: ignore
  comment if it is known in advance that 0D-to-scalar cast will be performed.

[charris]

Thanks Bas.