Don't return pointers to static objects with return_value_policy::take_ownership.

[MaartenBaert]

This fixes -Wfree-nonheap-object warnings produced by GCC.

Description

When compiling the pybind tests with the latest GCC, the following warnings are produced:

In function ‘cast’,
    inlined from ‘operator()’ at pybind11/include/pybind11/pybind11.h:249:33,
    inlined from ‘_FUN’ at pybind11/include/pybind11/pybind11.h:224:21:
pybind11/include/pybind11/detail/../cast.h:653:13: warning: ‘operator delete’ called on unallocated object ‘int_string_pair’ [-Wfree-nonheap-object]
  653 |             delete src;
      |             ^
pybind11/include/pybind11/detail/../cast.h: In function ‘_FUN’:
pybind11/tests/test_builtin_casters.cpp:269:40: note: declared here
  269 |     static std::pair<int, std::string> int_string_pair{2, "items"};
      |                                        ^
In function ‘cast’,
    inlined from ‘operator()’ at pybind11/include/pybind11/pybind11.h:249:33,
    inlined from ‘_FUN’ at pybind11/include/pybind11/pybind11.h:224:21:
pybind11/include/pybind11/stl.h:190:5: warning: ‘operator delete’ called on unallocated object ‘lvv’ [-Wfree-nonheap-object]
  190 |     PYBIND11_TYPE_CASTER(Type, const_name("List[") + value_conv::name + const_name("]"));
      |     ^
pybind11/include/pybind11/stl.h: In function ‘_FUN’:
pybind11/tests/test_stl.cpp:179: note: declared here
  179 |     static std::vector<RValueCaster> lvv{2};
      | 

Apparently two test cases were returning pointers to static objects with return_value_policy::take_ownership, which causes pybind to delete these objects later. This commit fixes that by allocating a copy and returning it instead of a pointer to the original.

Suggested changelog entry:

Fix ``-Wfree-nonheap-object`` warnings produced by GCC by avoiding returning pointers to static objects with ``return_value_policy::take_ownership``.

[Skylion007]

We still need to check that pointers to STL containers are castable.
Nvm, forgot new returns a pointer. Thoughts @rwgk ?

[Skylion007]

@henryiii @rwgk Can we get a 2nd pair of eyes?

[rwgk]

We still need to check that pointers to STL containers are castable. Nvm, forgot new returns a pointer. Thoughts @rwgk ?

First thought: Huh? Why didn't ASAN and valgrind catch this? Pretty surprising, glad you discovered this.

Second thought: py::return_value_policy::reference would seem like a better fix. Is there a catch with that?

[Skylion007]

Yeah, I agree that changing the return_value_policy seems better.