When a handler is added to an informer we must deliver to it
all existing objects of its type to match the Shared Informer
behavior.

For queued informers we spread the adds out over an array of
initial-add-specific queues, so that we don't dump a whole set
of objects into the queues that other informers are also
receiving events from; only the handler being added should
get this set of initial objects.

Unfortunately the initial add logic to spread objects across
the initial-add-specific queues took its data from the regular
queues, which caused initial add queues to become unbalanced
when the regular queues were processing lots of objects.