Porting/adapting Dustin's PR #2839 to smart_holder branch

[rwgk]

Hi @virtuald, this PR ports your PR #2839 to the smart_holder branch, ready to use if you like to give it a try. The branch is a strict superset of current master, with pybind11/smart_holder.h as an optional-use add-on.

Some remarks:

• The test you're adding under Attach python lifetime to shared_ptr passed to C++ #2839 was copied over to here, with a different name: test_class_sh_trampoline_shared_ptr_cpp_arg. This way we can merge your PR into master, and then into smart_holder without conflicts. After the merge I will try to consolidate the copies, to remove the duplication.

• The copy-modified test doubles as an example for how to use smart_holder (include smart_holder.h, add PYBIND11_SMART_HOLDER_TYPE_CASTERS macros, change py::class_ to py::classh).

• There is one section in the copy-modified test that expects different behavior compared to your original test, prominently marked with comments. I'm currently unclear if the has_alias/is_alias distinction is actually desirable. I tried a little bit to emulate it, but gave up when I couldn't find a straightforward way to make it work. If you feel strongly the distinction is important, I could try harder to make it work also in the smart_holder code.

• This PR (and the entire smart_holder branch) does NOT make any changes that would require a PYBIND11_INTERNALS_VERSION increment. The trick I'm using is to take a free bit in the smart_holder type, to track pointee_depends_on_holder_owner. Currently the smarter_holder type is not under any version lock policy, because it's still a brand new addition, but after your PR is merged on master, I will look into using your has_alias/is_alias bits instead. I think it makes more sense to keep those bits separate from the rest of the smart_holder logic.

• Your shared_ptr_deleter appears here as shared_ptr_dec_ref_deleter (I have two other deleters, therefore the more specific name). Note that it works with void operator()(void *), the T isn't actually needed. A small simplification / minor optimization, also makes it easier to reuse between our implementations maybe if we want to make things pretty down the road.

• Note that the smart_holder also support passing unique_ptr back to C++ if safe (disowning the Python object), but if has_alias is true this feature is disabled; a ValueError is raised.

• For the record, I tried out my idea outlined previously under Attach python lifetime to shared_ptr passed to C++ #2839: populating the smart_holder immediately with a shared_ptr that has your deleter. It was a bad idea: makes a nice reference cycle right there, undetectable by the gc. Building a new shared_ptr each time one is needed for passing to C++ is the the only way I could find to achieve the desired behavior.

[rwgk]

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