Snapshot caching, introduced in 623a9ba7, did not increment
xactCompletionCount during subtransaction abort. That could lead to an older
snapshot being reused. That is, at least as far as I can see, not a
correctness issue (for MVCC snapshots there's no difference between "in
progress" and "aborted"). The only difference between the old and new
snapshots would be a newer ->xmax.

While HeapTupleSatisfiesMVCC makes the same visibility determination, reusing
the old snapshot leads HeapTupleSatisfiesMVCC to not set
HEAP_XMIN_INVALID. Which subsequently causes the kill_prior_tuple optimization
to not kick in (via HeapTupleIsSurelyDead() returning false). The performance
effects of doing the same index-lookups over and over again is how the issue
was discovered...

Fix the issue by incrementing xactCompletionCount in
XidCacheRemoveRunningXids. It already acquires ProcArrayLock exclusively,
making that an easy proposition.

Add a test to ensure that kill_prior_tuple prevents index growth when it
involves aborted subtransaction of the current transaction.

Author: Andres Freund
Discussion: https://postgr.es/m/20210406043521.lopeo7bbigad3n6t@alap3.anarazel.de
Discussion: https://postgr.es/m/20210317055718.v6qs3ltzrformqoa%40alap3.anarazel.de