Revamp the PIPE test a bit. Use a calibration loop to make the test always

Revamp the PIPE test a bit.  Use a calibration loop to make the test always
run for a consistent period of time (around 6 seconds) no matter what the
specified block size is.  Allow a third 'label' argument to be specified
and reformat the MB/sec output to include the label and block size
to make it easier to pick off results for gnuplot.  Finally, use mmap()
instead of malloc() to create a cache-friendly read and write buffer to
reduce noise in the performance results.