[BUG] Weird issue only on manylinux1 with numpy

Dumping this here since I need to fix this and move on for now (it's an issue in Awkward). Only on manylinux1 (2010 is fine), the following code produces a really weird result:

py::print(typeid(obj.ptr()).name());
py::print(typeid(py::module::import("numpy").attr("newaxis").ptr()).name());
py::print(typeid(Py_None).name());
py::print((long long) obj.ptr());
py::print((long long) py::module::import("numpy").attr("newaxis").ptr());
py::print((long long) Py_None);
py::print(obj.is(py::module::import("numpy").attr("newaxis")));
py::print(obj.ptr() == py::module::import("numpy").attr("newaxis").ptr());

Prints:

P7_object
P7_object
P7_object
8098416
8098416
8098416
False
True

Why is obj.is(py::module::import("numpy").attr("newaxis")) False? It's the same memory location as Py_None, but somehow is not "is" identical, only here in pybind11.

PS: obj is None, numpy.newaxis is None.

Run the supplied C++ reproducer with numpy under manylinux1 and manylinux2010, comparing the reported pointer values and identity results. Then inspect pybind11's implementation of obj.is to trace the platform-specific behavior. Done means identifying the cause and verifying a fix or documented resolution for the None and numpy.newaxis identity comparison.