gh-116126: Implement PEP 696

[JelleZijlstra]

This implements both the grammar and compiler changes and changes to typing.py behavior (the latter copied from typing-extensions).

• Issue: Implement PEP 696 (Type parameter defaults) #116126

[JelleZijlstra]

This should be ready for review now, I'll take it out of draft once the tests pass. Notes:

• This PR intentionally does not cover documentation or typing.py behavior; I'll leave that to another PR.
• The PEP 695 test suite helped me find some very interesting issues with the symtable effects of this change. I think the current solution should work, but definitely didn't expect that to be the trickiest aspect of the change.
• It's unfortunate that the new AST field is called default_ in Python. It can't be default in C because that's a keyword. I may need to change the AST generation machinery to support aliasing a field so the Python-visible name can be just "default".

[pablogsal]

Will review this weekend

[JelleZijlstra]

@pablogsal have you had a chance to look at this? The grammar changes should be quite straightforward, so hopefully it's not complicated to review.

I'll spend some time today trying to make it so the field is called "default" not "default_" in Python, because having "default_" in the Python-visible APIs would be ugly.

[AlexWaygood]

I would say "is not the typing.NoDefault singleton" rather than "is not equal to typing.NoDefault", since we usually argue that an identity check is more idiomatic for singleton objects such as None, NotImplemented and Ellipsis

[AlexWaygood]

Suggested change

	>>> T.__default__ is typing.NoDefault
	True
	>>> S = TypeVar("S", default=None)
	>>> S.__default__ isNone
	True
	>>> T.has_default()
	False
	>>> T.__default__ is typing.NoDefault
	True
	>>> S = TypeVar("S", default=None)
	>>> S.has_default()
	True
	>>> S.__default__ isNone
	True

[JelleZijlstra]

I thought of that but it felt out of place in the docs for NoDefault, since the cases you added don't use NoDefault at all.

[AlexWaygood]

Fair enough -- I thought it was quite nice to see together in one example how the two concepts interrelate, but I definitely don't feel strongly!

[JelleZijlstra]

@ericsnowcurrently the C globals check is failing because this PR now adds a new static type and global singleton, typing.NoDefault (closely analogous to builtins.NotImplemented). What do you think we should do? We could simply allowlist the new globals, or somehow make NoDefault not rely on C globals—but the latter feels like it would be a lot more complicated.

[ericsnowcurrently]

For now let's just whitelist the new singleton, as long as it is stateless and immortal. We can circle back later if the whitelist is a problem.

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @JelleZijlstra for commit 5f6fdfd 🤖

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