gh-101766: Fix refleak for _BlockingOnManager resources from test suite level

[corona10]

• Issue: Refleak in test_importlib on aarch64 RHEL8 #101766

[corona10]

@exarkun @gvanrossum @sobolevn

• I revert the 775f881 since @exarkun pointed out that the change will cause the regression which was already resolved.
  see: gh-101766: Fix refleak for _BlockingOnManager resources #101942 (comment)
• Rather than I tried to clean up the global dictionary at the test suite level. It will reduce the possibility of leakage at the test level. (Not 100% but it doesn't occur under 10 times of execution. but without this patch, the leak was reproduced within 10 times of execution)

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @corona10 for commit a85f0e6 🤖

If you want to schedule another build, you need to add the 🔨 test-with-refleak-buildbots label again.

[sobolevn]

Wow, this turned out way more complex than I initially thought! 👍

[gvanrossum]

This works too. Many tests do something like this. Is there standard infrastructure for clearing caches?

[corona10]

This works too. Many tests do something like this. Is there standard infrastructure for clearing caches?

I am not sure, let's create a issue for that. We will need it.