Resolve reference counting error in the RNA API which mixed up ownership
between ExtensionRNA::data & StructRNA::py_data.
In practice this generally worked as RNA_struct_free_extension behaved
as if ExtensionRNA owned the reference instead of StructRNA, so as long
as only one reference was removed, there wasn't any difference.
When re-registering Python classes with RNA, ownership wasn't handled
properly which could result in negative reference counts,
crashing in some cases.
bfb6b184f5