New semantic analyzer: Fix --disallow-untyped-defs with attrs plugin and forward references

[ilevkivskyi]

The fix essentially mirrors the one for dataclasses. Note I temporary add final_iteration to the old analyzer, because plugin files import everything from the old one.

The non-trivial part is removing previously generated functions in plugins to make them re-entrant.

[ilevkivskyi]

Actually, it looks like the guard against final iteration is not necessary, I am going to remove it for now.

[ilevkivskyi]

Oh, it is actually needed, I just run the tests without NEWSEMANAL=1.

[ilevkivskyi]

The test failures are actually hard to fix. They highlight a problem with all plugins that use add_method() -- they are not re-entrant. The add_method() calls .append() on class body that causes multiple functions with same name added if there is any deferral in the file. I am surprised no tests failed because of this before.

[msullivan]

This is solid for a bugfix but I think we need to consider if there is a cleaner way to allow plugins to handle these issues. Explicitly checking for the new analyzer and explicitly checking final_iteration are pretty nitty-gritty details that we don't want to have to inflict everywhere.

[ilevkivskyi]

Explicitly checking for the new analyzer and explicitly checking final_iteration are pretty nitty-gritty details that we don't want to have to inflict everywhere.

Yes. One of this will go after the new one is the default. For the second one, possible option is to have defer_if_possible(). It will be actually useful in general, I think we already use this pattern in several places in semantic analyzer. I will create an issue for this.

[ilevkivskyi]

@msullivan Opened #6725