[mypyc] Fix __init_subclass__ running before ClassVar instantiations

[VaggelisD]

Fixes mypyc/mypyc#1086

We also ran into this in SQLGlot; To compile our AST we swapped any metaclasses for __init_subclass__ in which we dynamically set some attributes in it like so:

  class Expression:                                                                                                                                                                                                                                  
      arg_types: ClassVar[...] = {}                                                                                                                                                                                                      
                                                            
      def __init_subclass__(cls, **kwargs: object) -> None:                                                                                                                                                                                          
          cls.required_args = {k for k, v in cls.arg_types.items() if v}

  class Select(Expression):
      arg_types: ClassVar[...] = {...}

I introduced the InitSubclass primitive to allow for fine-grained control of when this gets called and moved it from ExtClassBuilder::__init__ (i.e allocate_class > CPyType_FromTemplate) to ExtClassBuilder::finalize

[tobymao]

@JukkaL thanks for the quick reviews so far, any chance we can get this reviewed as well so that we can unblock sqlglots compilation with mypyc?

this along with #20917, we'll soon have a new PR that fixes another vtable bug that sets getters and setters to the same slot

[JukkaL]

Thanks, looks good. I have just one minor code quality suggestion.

[JukkaL]

Can you remove the | None part here, which would let you skip the is not None check below? It looks like this is always not-None.

[VaggelisD]

Thanks for the review, done! Thought I had went over exactly that and deemed None necessary, not sure why..