gettime: for better accuracy calculate cycles_per_msec instead of cycles_per_usec

fio (before the nsec changes) would finish about 9 sec early for a 16 hour job
when relying on the CPU clock. This is because fio calculates cycles_per_usec
to carry out the clock ticks to time conversion. cycles_per_usec only provides
4 significant digits. Changing this to cycles_per_msec provides 7 significant
digits and makes the actual job run time more closely match wall time.