Buffer must always either contain NULL or valid references
(assuming the dtype supports references) in order to allow
cleanup in case of errors.

It is thus unnecessary to clear buffers before every copy, if
they contain NULL, all is fine. If they contain non-NULL, we should
also DECREF those references (so it would be incorrect as well).

Buffers thus need the memset exactly once: directly upon allcoation
(which we do now).  After this any transfer from/to the buffer needs
to ensure that the buffer is always in a good state.

This is necessary because in some rare cases (reductions), we may
not actually use the full buffer.  In that case, the cleanup-on-error
code would have to grow smart enough to handle these cases.
It seems much simpler to just always initialize the full buffers, even
if we may not end up using them.

Admittedly, the old logic might have skipped the buffer clearing
(especially the full buffer) in a fair bit of cases, but since
this is only relevant for `object` dtype, I assume this is fine.

BUG: for MINGW,  threads.h existence test requires GLIBC > 2.12

BLD: remove unnecessary flag `-faltivec` on macOS

MAINT: Allow more recursion depth for scalar tests.

  Supporting PowerPC/AltiVec on macOS is no longer necessary,
  even if the Mac/G5 is still running having `-faltivec` or
  replacing it with the new AltiVec flag `-maltivec` wouldn't
  increase that much performance without raw SIMD.

  note: the flag was enabled on non-intel platforms which
  causes fatal build errors on macOS/arm64.

The test_operator_object_left and test_operator_object_right tests in
test_scalarmath.py set a maximum recursion depth of 100 that caused an
error in gitpod because it was too close to the starting depth of 75.
The fix here is to raise the maximum depth to 200.

Closes #18749