The logic here used to consist of a couple of ad-hoc checks,
followed by a general assumption that if something had already
been emitted, it could be referenced directly, whereas everything
else had to go through a GOT entry.
This is way too conservative. Instead, let's try to correctly
calculate what translation unit an entity is going to end up in.
Previously even if a type's metadata was optimized away, we would still
emit a field descriptor, which in turn could reference nominal type
descriptors for other types via symbolic references, etc.
