Enable defining custom __new__

[Skylion007]

Description

• This solves the bugs preventing us from defining our own new methods on classes that was brought up in [BUG]: Overriding __new__ (or type-level __call__) for a pybind11 class should be supported #3253 .

Suggested changelog entry:

* Enable defining custom ``__new__`` methods on classes by fixing bug preventing overriding methods if they have non-pybind11 siblings.

[rwgk]

How did I arrive at 194033e?

When testing in the Google environment, I ran into this assert:

https://github.com/python/cpython/blob/d5650a1738fe34f6e1db4af5f4c4edb7cae90a36/Objects/typeobject.c#L3117

The error wasn't obvious at all, therefore I inserted this right before that assert:

    if (PyErr_Occurred()) {                                                     
      PyErr_Print();                                                            
      assert(0);                                                                
    } 

That way I got to see the FutureWarning. After disabling it with the unconditional #define, all tests ran ASAN clean.

While looking around, I saw that all new-style "__init__" return a pointer (from a placement new). Therefore I changed the new code accordingly, before I discovered the FutureWarning. I didn't go back to test if it also works when returning a reference (I believe returning a pointer is preferable regardless, although I don't have deep insights).

That's all I know. 🙂

[Skylion007]

Defining a custom new function for enums could help close #1177. Also we are putting a class method in the tests now so we might want to add a class method shim as mentioned in #1693 now to py_class as the current way of defining class methods is bulky and not very intuitive.

We don't seem to have a way to check the py::type of a the C++ class for the first argument though, which is a bit annoying.

[Skylion007]

This syntax is at the very least not terrible until we have proper classmethod support.

[Skylion007]

I actually just realized those PTRs in test should probably be ref since they can't be null. Probably can go back and fix them when we write the actual documentation for this feature at some point @henryiii along with proper classmethod support perhaps

[Skylion007]

I also wish there was a cleaner way to not have to implement __init__ if new is defined, but I can't figure out how to hack to the metaclass not to complain.

[rwgk]

Maybe something like .def_new(...) or .def(py::new(...)) that wraps the two methods you have right now?