opt: remove some generic code bloat

[davidhewitt]

Comprised of three commits when just playing around this morning:

• Use fn instead of closure in LazyStaticType::get_or_init to avoid creating a new monomorphization of GILOnceCell::get_or_init for each PyClass
• Use fn() instead of impl FnOnce in extract_argument default / from_py_with arguments, reducing number of monomorphizations of those functions. I think this is more correct anyway (we don't need to support capturing closures).
• Change #[pyclass] machinery to create an iterator of PyClassItems instead of relying on a "visitor" closure. I think it's tidier, and again reduces the amount of code that needs to be generated per PyClass.

Overall this reduced llvm lines by about 2% in the pytests crate.