Now we take the difference of the time to process the input
and the time to run with no input. This compensates for
slow startup time in dynamic languages.
See comments on 2dcef8a
'make bench' runs the program 20 times against a file
composed of 10 copies of Pro Git (about 5MB in all).
It then does statistics on the results.
We run it with high priority to get more consistent timings.
