Fixed two bugs introduced in the previous commit (hash 4319853)

The first bug is that packet buffers arriving out-of-order are not returned to
the driver properly; this will exhaust the driver's rx buffers eventually. The
second bug is some experimental junk left in the homa benchmark code.

Some people never thought this message would appear in the git log.

P.S. We now consider RAMCloud to be usable for (some) real applications.

Print log message in TableManager.cc when tables are deleted.

Modifed Metrics to handle cleanly the case where multiple services

are recovering on the same server (clear statistics only when the
first service starts recovery, output statistics only when the last
service finishes).  This cleans up the server code quite a bit, but
makes scripts/metrics.py a bit uglier.

This is the code used for the CFP submission for our SOSP 2011 paper.