copy/move rvp does not work as expected for already registered python objects

The copy/move return value policy does not apply correctly to a C++ reference returned from the function in case the reference being returned is already associated with a previously registered pybind11 instance. The underlying problem is that the cast function in type_caster_generic class returns an existing instance, if any, before even considering the intended rvp. To fix this issue, the search for existing instances needs to be predicated on the rvp not being a copy/move. The following code reproduces the problem.

// C++ binding
#include "pybind11/pybind11.h"
namespace py = pybind11;

struct A { 
  int x;
  A(int x) : x(x) {}
  A(const A& a): x(a.x * 100) {}
  void m() { printf("A.m() = %p, %d\n", this, x); }
};

A& foo(A& a) {
  return a;
}

PYBIND11_MODULE(example, m) {
  m.doc() = "pybind11 example plugin";
  py::class_<A>(m, "A")
  .def(py::init<int>())
  .def(py::init<const A&>())
  .def("m", &A::m);
  m.def("foo", foo, py::return_value_policy::copy);
}

Python code

from example import A, foo, bar
a = A(1)
a.m()
foo(a).m()

Possible output

A.m() = 0x7052ab, 1
A.m() = 0x7052ab, 1

Start at the type_caster_generic cast function and reproduce the behavior with the C++ binding and Python example in the issue. Trace how existing registered instances are selected for the return value policy, then add regression coverage showing that copy/move policies produce the intended result and run the relevant test suite.