A few thoughts:

1. The merged PR releases GIL for any Python code run in `with gc.ensure_disabled()`.  This is just plain wrong.

2. The merged PR crashes on debug build of CPython.

3. I don't actually understand this feature.  Our GC is not per OS thread, which means that inside `with gc.ensure_disabled()` the GC can become suddenly **enabled**, if another thread enables it.  This API is just misleading (maybe the name of the new context manager is bad—it cannot ensure anything).  There's even a unittest that tests this!

4. This new API can be trivially implemented in pure Python:

   @contextmanager
   def disable_gc():
       gc.disable()
       try:
           yield
       finally:
           gc.enable()

5. While such pure Python version shares the problem discussed in (3), the currently committed C implementation doesn't fix them either.

IMO this entire issue needs to be properly debated on python-ideas first and the PR should be reverted from beta-1.  If the latter is not possible, OK, let's revert it for beta-2 and for 3.8.