gh-148680: Replace internal names with type_reprs of objects in string representations of ForwardRef

[DavidCEllis]

This adds a new ForwardRef.__resolved_forward_str__ attribute which takes the __forward_arg__ and replaces any internal __annotationlib_name_x__ names with the type_repr of the object they represent.

Before:

ForwardRef('ref | __annotationlib_name_1__', is_class=True, owner=<class '__main__.Example'>)
ref | __annotationlib_name_1__

After:

ForwardRef('ref | str', is_class=True, owner=<class '__main__.Example'>)
ref | str

This approach replaces the extra_names with their type_repr when the forwardref is evaluated to a string. I'm fairly sure that trying to get the original names as they are in the source would be both slower and significantly more complicated.

This logic is largely based on something I already have for my reannotate library.

• Issue: annotationlib - evaluating forward references in STRING format can 'leak' internal names #148680

[JelleZijlstra]

We might need to cache this, it's probably pretty slow. What do you think?

[DavidCEllis]

This is in fact, not very fast. Definitely worth caching if it's going to be accessed frequently.

[DavidCEllis]

>>> a_anno
ForwardRef('unknown | str | int | list[str] | tuple[int, ...]', is_class=True, owner=<class '__main__.Example'>)
>>> b_anno
ForwardRef('unknown', is_class=True, owner=<class '__main__.Example'>)

>>> a = timeit(lambda: a_anno.__resolved_forward_str__, number=10_000)
>>> b = timeit(lambda: b_anno.__resolved_forward_str__, number=10_000)
>>> a
0.18359258100008446
>>> b
0.0018204959997092374
>>> a / b
100.8475607907994

[DavidCEllis]

I've added a cache and ended up shortening the name of the property and the cache. Adding the cache did mean adding an extra slot to both classes.

Not completely sold on the names I have if you have something better.