Speed up pruning of the threading tree

Previously, a lot of time was spent on looking up many items in an
internal hash table, and additional time was wasted by renumbering the
siblings all the time in case multiple removals took place at the same
time.

While each EXPUNGE is still being processed on its own, the actual
removal of each of these items now happens much faster (roughly 30% on a
synthetic benchmark based upon the callgrind's instruction fetch
counts).

Change-Id: I4fa6e0a884859d0105fc535c70ca1d0fb3c57d5b