Using dynamic_cast<AliasType> to determine pointee_depends_on_holder_owner.

[rwgk]

This PR is a follow-on to PR #2902, implementing a refinement of a safety guard that previously prevented not only unsafe behavior, but also safe behavior. All relevant situations are exercised by the newly added test_class_sh_virtual_py_cpp_mix.py. Specifically, e.g. m.get_from_cpp_unique_ptr(m.Base()) fails before this PR.

The critical change in this PR is in smart_holder_type_casters.h:

-        v_h.holder<holder_type>().pointee_depends_on_holder_owner = has_alias;
+        v_h.holder<holder_type>().pointee_depends_on_holder_owner
+        = dynamic_raw_ptr_cast_if_possible<AliasType>(v_h.value_ptr<WrappedType>()) != nullptr;

has_alias is true for any instance of a py::class_ with an alias class (aka trampoline). E.g. it is true for m.Base() and PyBase() in test_class_sh_virtual_py_cpp_mix.py. The newly used dynamic_cast-based test is true only for PyBase().

The idea for using a dynamic_cast here goes back to:

pybind11/include/pybind11/detail/init.h

Line 49 in 0e01c24

return dynamic_cast<Alias<Class> *>(ptr) != nullptr;

The exception message for the safety guard (also in smart_holder_type_casters.h) was made more precise and helpful:

         if (self_life_support == nullptr && holder().pointee_depends_on_holder_owner) {
-            throw value_error("Ownership of instance with virtual overrides in Python cannot be "
-                             "transferred to C++.");
+            throw value_error("Alias class (also known as trampoline) does not inherit from "
+                              "py::virtual_overrider_self_life_support, therefore the "
+                              "ownership of this instance cannot safely be transferred to C++.");
         }

Not directly related to the behavior change, py::detail::virtual_overrider_self_life_support was moved to py::virtual_overrider_self_life_support, to reflect that it is part of the public interface.

The behavior change under this PR has an effect on test_class_sh_trampoline_shared_ptr_cpp_arg.py, which was adopted from PR #2839 previously. test_shared_ptr_alias_nonpython is now in line with the behavior under PR #2839, but test_shared_ptr_arg_identity behaves differently here. This may need another look, but it appears that the has_alias/is_alias logic under PR #2839 may need a refinement similar to what is implemented in this PR.

[rwgk]

The CI is green, also with PYBIND11_USE_SMART_HOLDER_AS_DEFAULT (#2879). Google-internal testing is also green. Merging here to unblock follow-on PyCLIF work. This change will get reviewed internally. I'll open a new PR if there are requests for changes.