Refactor stubgen to do semantic analysis

[dmoisset]

Currently stubgen just builds an AST and works mostly at a syntactic level. Given that, most type annotations during its executions are actually instances of UnboundType and forces stubgen to make some assumptions because of lack of information (for example in a type X.Y.Z, which part of that path is a module name, or if actually X = Y is a type alias or not)

This comes from #3169 (comment)

[ilevkivskyi]

Thus is now supported (and controlled by a flag, see https://mypy.readthedocs.io/en/latest/stubgen.html#specifying-how-to-generate-stubs)