Fix crash in function with unpack and constrained TypeVar

[ilevkivskyi]

Fixes #21907

This fixes the crash by not normalizing callables too soon (note the TypedDict test passes on master, but I still add it to prevent similar regression there).

cc @JukkaL

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[willy-b]

I can confirm this change resolves the issue reported in #21907 .

In my proposed fix for #21907
in willy-b@eb59611 from 2026-09-17 (2026-09-16 PT),

I also found it sufficient to avoid calling expanded.with_normalized_var_args() here
downstream of checker.py expand_typevars, also adding a optional parameter to expand_type as you do here setting a flag on ExpandTypeVisitor controlling whether to normalize here.

However I was divided between stopping the with_normalized_var_args call directly here as you do or stopping needs_normalization from being set above at:

if var_arg is not None and isinstance(var_arg.typ, UnpackType):
            needs_normalization = True
            arg_types = self.interpolate_args_for_unpack(t, var_arg.typ)
        else:
            arg_types = self.expand_types(t.arg_types)

I had not confirmed where interpolate_args_for_unpack was being used to determine whether that should still be run or not in this case (expanded.with_normalized_var_args() is sufficient to fix this one issue but I was worried about side effects of that or this approach). Thanks