References from autodoc signatures should be resolved via `py:obj` role, not `py:class`

I have a code base that uses lots of type aliases. Most of them are documented as py:type, and some literal types are documented as py:data. Because of this, references to these objects are not resolved, because python domain uses py:class role to search for them.

I believe that we should use py:obj for all references generated in signatures. I don't see anything that can break if we change current behavior. After all, Python domain doesn't handle objects of different types sharing the same fully-qualified name, so there should be no collisions.

Alternatively, we can make a new role for this (I use lua:_auto in my Lua extension). This will allow easily suppressing nitpicks originating from signatures.

Trace how autodoc signatures generate references and how the Python domain resolves the py:class and py:obj roles. Compare the proposed behavior with the linked Lua _auto implementation; done means documented type aliases and literal types resolve from signatures without breaking existing reference behavior.