switch to using new casting as default to run full CI

Initialize casting to silence compile warnings only...

BUG: As my comment even said, out_needs_api is caller-initialized

so I shouldn't init it to 0, it might disable api again.

Simplify tests again, it seems unnecessar to go this complex

Simplify `#if` as the variable should always be defined

TST: Add more tests (and some related fixes)

This fixes a few small things (i.e. the decref loop can be NULL
and checking for writeable was incorrect in the private API).

Mostly adds new tests for some internal API as well as for the copy
of the auxiliar data. The nditer test however also is a heavyweight
tests for the general multistep casting

It seems the only reason this did not cause problems before was
that the test checking for C warnings only ran on a run where
these were always 0 and thus undefined (and so were not redefined
later).
The `#ifndef` would have to happen at a later time apparently, so
just avoid it.

The second patch in a series of pull-requests aims to facilitate the migration
  process to our new SIMD interface(NPYV).

  It is basically a process that focuses on getting rid of the main umath SIMD source `simd.inc`,
  which contains almost all SIMD kernels, by splitting it into several dispatch-able sources without
  changing the base code, which facilitates the review process during the move to NPYV(universal intrinsics).

  In this patch, we have moved the the following raw SIMD loops to the new dispatcher:
   - FLOAT_exp,   DOUBLE_exp
   - FLOAT_log,   DOUBLE_log
   - FLOAT_frexp, DOUBLE_frexp
   - FLOAT_ldexp, DOUBLE_ldexp

   The new code improves the performance of non-contiguous memory access
   for the output array without any reduction in performance.
   For PPC64LE the performance increased by 2-3.0, and 1.5-2.0 on aarch64.

   The first patch in a series of pull-requests aims to facilitate the migration
   process to our new SIMD interface(NPYV).

   It is basically a process that focuses on getting rid of the main umath SIMD source `simd.inc`,
   which contains almost all SIMD kernels, by splitting it into several dispatch-able sources without
   changing the base code, which facilitates the review process in order to speed up access to the nominal target.

   In this patch, we have moved the arithmetic operations of real and complex for single/double precision
   to the new CPU dispatcher.

   NOTE: previously, the SIMD code of AVX2 and AVX512F for single/double precision wasn't dispatched in runtime before.

Lets defer further touch ups to later... One more run, since the
last one errored (hopefully due to an old failure not merged
correctly)

   - only covers sqrt, absolute, square and reciprocal
   - fix SIMD memory overlap check for aliasing(same ptr & stride)
   - unify fp/domain errors for both scalars and vectors

 - use plain variables
 - clean up aligned allocate
 - use `PyArg_ParseTuple` for empty args
 - use `Py_ssize_t` instead of `unsigned` and `size_t`
 - improve coding style
 - no need for a custom raises assertions
 - use parametrize instead of inner loops
 - leave a comment about nature of mode testing unit
 - shift to get max/min of int72
 - add more info to repr of vector object
 - get ride of exec() and use type() instead
 - use `.inc` as extension for sub-headers instead of `.h`
 - add `FMA4` and drop `SSE41` from _SIMD targets

Co-authored-by: Eric Wieser <wieser.eric@gmail.com>