The number of indices is limited to 2*MAXDIMS currently to allow
mixing integer indices, e.g. with new indices `np.newaxis` (one
removes output dimensions, the other adds new ones).
This means that more than MAXDIMS advanced indices can be passed
on to the advanced indexing machinery (`MapIterNew`), which did
not check for this possibility.

Closes gh-18145

Closes (the later point) in gh-17965 and reverts parts of
gh-17817.
Shapely did rely on being able to raise a NotImplementedError
which then got ignored in the attribute lookup.
Arguably, this should probably just raise an AttributeError to
achieve that behaviour, but it means we can't just rip the band-aid
off here.

Since 1.20 is practically released, just reverting most of the change
(leaving only recursion and memory error which are both arguably
pretty fatal).
Ignoring most errors should be deprecated (and I am happy to do so),
but it is not important enough for 1.20 or very important in itself.

Closes gh-17965

After an unrelated fix, the logic for string and object promotion
was incorrect briefly, this fixes it to be correct (symmetric).
Before, string and unicode would return that
`string.__common_dtype(object)` is in fact `string`, which is of
course incorrect. (since `object.__common_dtype__(other)` always
returns `object` this depended on the order, and the NumPy tests
apparently did only test the opposite direction (or nothing).

  It also fixes conversion warning between `__mmask16` and `__mmask8`
  on msvc2019 when logical intrinsics of AVX512DQ are available.

I somehow managed to miss that half breaks the order of dtypes and
has a higher number than the strings. Could be backported, but it
doesn't really matter, since it only makes a difference if the
compile time flag is used and even then is pretty fringe.

Co-authored-by: Matti Picus <matti.picus@gmail.com>
Co-authored-by: h-vetinari <h.vetinari@gmx.com>
Co-authored-by: Derek Homeier <dhomeie@gwdg.de>

  The new path becomes `build/src.*/numpy/distutils/include/npy_cpu_dispatch_config.h`
  instead of `numpy/core/src/common/_cpu_dispatch.h`.

  The new path allows other projects to re-use the CPU dispatcher
  once we decide to expose the following headers:
    - `numpy/core/src/common/npy_cpu_dispatch.h`
    - `numpy/core/src/common/npy_cpu_features.h`

Previously, the dtype was used, this now assumes that we want to
cast to a string of (unknown) length.  This is a simplified version
of what happens in `np.array()` or `arr.astype()` (it does never
inspect the values, e.g. for object casts).

This is more complex as I would like, and with the refactor of
ResultType and similar can be cleaned up a bit more hopefully.

Note that currently, object to "S" or "U" casts simply return
length 64 strings, but with the new version, this will be an error
(although the error message probably needs improvement).
This is a behaviour inherited from other places however.