Actually, the ideal thing to happen would be for PL/Python to somehow become a shared library that provides the typecasting routines, but I'm afraid that's a very big task and difficult to get accepted upstream. It's sad that PL/Python already includes all the code you need to transform Python lists to Postgres arrays, bools to PG bools and so on and every project like this one has to duplicate all that logic.
BTW: in my approach I opted for passing the entire Qual node as a string (using nodeToString) and my plan was to write a Python parser for that format that takes a string and returns a Python object. Not sure how hard that will turn out.