The new nleft_dead_tuples and nleft_dead_itemids fields are confusing
and do not seem like the correct way forward.  One of them is tested
via an assertion that can fail, as it has already done on buildfarm
member topminnow.  Remove the assertion and the fields.

Change the logic for the case where a tuple is not initially pruned
by heap_page_prune but later diagnosed HEAPTUPLE_DEAD by
HeapTupleSatisfiesVacuum.  Previously, tupgone = true was set in
that case, which leads to treating the tuple as one that will be
removed.  In a normal vacuum, that's OK, because we'll remove
index entries for it and then the second heap pass will remove the
tuple itself, but when index cleanup is disabled, those things
don't happen, so we must instead treat it as a recently-dead
tuple that we have voluntarily chosen to keep.

Report and analysis by Tom Lane.  This patch loosely based on one
from Masahiko Sawada, but I changed most of it.