The problem was that once a container got a parent, it did not change it anymore
due to the child_elligible condition, but the parent might have been assigned
from an incomplete References sequence.
Now, we make sure the last reference gets to be the message's parent (following
the JWZ's algorithm), reparenting the message if necessary. This makes sense, as
the last parent-child relationship (between last ref and the message) is the
most reliable piece of info here.
Instead of child_elligible, we now only check that the new parent is not a
descendant of the current message, to prevent making a loop. Everything else is
fine, as it only moves a subtree around.
mu_container_append_siblings was showing up high in profiles as it has to
walk chains of next->next->next->... pointers to find the last one. we now
cache the last link in the chain. for listing ~ 23K messages, this saves
about 20%.