gh-151728: Partially revert "Clear the typing caches at interpreter shutdown (GH-154858)"

[encukou]

Unfortunately, adding the atexit function causes reference leak tests to fail.
Broken Tier 1 buildbots, revert or fix immediatelly. I currently don't have time to investigate.

This only reverts the atexit registration: it leaves the _clear_caches function, to simplify the regrtest utils code a bit.

cc @wjakob

• Issue: typing caches keep extension types alive past shutdown #151728

[wjakob]

@encukou Can you tell me where those leak test failures can be seen? The original PR afaik had no test failures.

[encukou]

Refleak tests take a long time, so they aren't run on GHA, but post-merge on Buildbot. You can see them here: https://buildbot.python.org/#/release_status

Or locally, using:

./python -m test test_typing -R:

(see ./python -m test --help for more on -R)

[JelleZijlstra]

Example failing build https://buildbot.python.org/#/builders/1610/builds/3766

test_typing leaked [9021, 9017, 9021] references, sum=27059
test_typing leaked [6278, 6276, 6278] memory blocks, sum=18832
Re-running test_typing in verbose mode
test_typing leaked [9021, 9017, 9021] references, sum=27059
test_typing leaked [6278, 6276, 6278] memory blocks, sum=18832

I'm guessing the issue is we reload typing.py and register the callback repeatedly? Though I'm surprised that would leak that much memory.