This is a follow-up to rBb44e6f2b3d32, for some reason that issue was
not detected back then: in some cases, DEG_iterator_objects_next() will
free the temp list of dupli objects once it does not need it anymore,
henceforth freeing the dupli_object_current memory of the DEGObjectIterData
that we are storing in the RNA_Depsgraph_Instances_Iterator struct.
And yes, the uglyness of that hack is getting even better now...
Found while trying to export dupliobjects with FBX...