Cleanup internals for garbage-collected `py::class_` objects

In case we'd forget, @EricCousineau-TRI made some good remarks on #2564, that we should still attend to:

Not sure if it'd be easily addressable, but there may be lingering references from other types here:

https://github.com/pybind/pybind11/blob/7c71dd386161703dc25860870f579cca0ff3a9a8/include/pybind11/detail/internals.h#L135-L137

e.g. you're freeing type A, but type B has a conversion related to that type

• implicit_casts looks like it's only used for MI, so by construction I don't think you'd have a dangling parent base.
• direct_conversions looks like it's only used for NumPy dtypes
• implicit_conversions looks like it requires that both types be alive, so also should be fine. However, you may have unbounded memory growth (e.g. multiple registrations for a re-registered C++ class)

So yeah, looks like everything is fine. Just may want to make a minor note on implicit_conversiosn being unbounded; however, it's unlikely that unbounded growth will matter much? (unless someone likes running the GC every 30sec for 24hrs?)

Originally posted by @EricCousineau-TRI in https://github.com/pybind/pybind11/pull/2564#discussion_r504003987

Start with include/pybind11/detail/internals.h at the referenced lines and read PR #2564, then trace how implicit_casts, direct_conversions, and implicit_conversions retain type references. Done means determining whether garbage-collected py::class_ objects can leave dangling or unbounded conversion entries and recording the justified cleanup or note.