Create a metaclass for each supplemental data type

[oremanj]

This allows many metaclass customizations without incurring the performance cost of allowing arbitrary metaclasses. As a side benefit, we get nontrivial construction and destruction of supplement types, and can check whether a particular nanobind type uses a particular supplement type. There are no additional per-instance performance or storage costs, no additional per-type storage costs, and negligible per-type performance costs.

The information about a particular supplement type is stored in a new nb::detail::supplement_data structure. An inline variable template is used to automatically create one of these in static data for each supplement type used in a particular extension module. Instances of nb_meta (i.e., metaclass types) gain a pointer to their supplement_data, which is null for the default metaclass nb_type_0. The non-default metaclass types nb_type_xyz are now stored in a nb_type_map_slow rather than a Python dictionary, and (except on free-threaded builds where everything is immortal) they are properly reference-counted so that they will be destroyed once all their instances are. This avoids leak warnings when assigning nanobind functions as their attributes.

Size cost:

• On my laptop this adds 781 bytes to a release build of libnanobind-static.a.
• One pointer per nanobind metatype, including the default one.
• One nb::detail::supplement_data structure (five pointers) stored in static data per supplement type used in an extension module.

I originally added some caching in static data to limit lookups by supplement type to one per supplement type per extension module. I removed this because it was a bit complex and I don't think it's necessary (we already do two type_map_slow lookups per nb_type_new, avoiding a third doesn't save all that much) but it would be easy to re-add. See commit 34739e3. The additional cost from that caching would be 104 bytes of code and two more pointers in supplement_data.

[oremanj]

The tests aren't running because the tests for a previous iteration hung. I don't have permission to stop the previous run https://github.com/wjakob/nanobind/actions/runs/13862493485. Maybe worth setting a timeout in github actions?