Change atexit behavior and PEP-489 multiphase init support.

When PyGILState_Ensure() is called in a non-Python thread before
PyEval_InitThreads(), only call PyEval_InitThreads() after calling
PyThreadState_New() to fix a crash.

Add an unit test in test_embed.

bpo-32096, bpo-30860:  Partially revert the commit
2ebc5ce4:

* Move structures back from Include/internal/mem.h to
  Objects/obmalloc.c
* Remove _PyObject_Initialize() and _PyMem_Initialize()
* Remove Include/internal/pymalloc.h
* Add test_capi.test_pre_initialization_api():
   Make sure that it's possible to call Py_DecodeLocale(), and then call
   Py_SetProgramName() with the decoded string, before Py_Initialize().

PyMem_RawMalloc() and Py_DecodeLocale() can be called again before
_PyRuntimeState_Init().
Co-Authored-By: Eric Snow <ericsnowcurrently@gmail.com>

* Remove Setup.config
* Always define WITH_THREAD for compatibility.

by increasing the number of iterations.  That appears sufficient to
expose the ref count problem fixed in this issue.
Patch suggested by Xiang Zhang

This new pre-initialization API allows embedding
applications like Blender to force a particular
encoding and error handler for the standard IO streams.

Also refactors Modules/_testembed.c to let us start
testing multiple embedding scenarios.

(Initial patch by Bastien Montagne)