This fixes a fundamental (and long-standing) bug in the sleep-average
estimator which is the root cause of the "contest process_load" problems
reported by Mike Galbraith and Andrew Morton, and which problem is
addressed by Mike's patch.

The bug is the following: the sleep_time code in activate_task()
over-estimates the true sleep time by 0.5 jiffies on average (0.5 msecs
on recent 2.5 kernels).  Furthermore, for highly context-switch
intensive and CPU-intensive workloads it means a constant 1 jiffy
over-estimation.  This turns the balance of giving and removing ticks
and nils the effect of the CPU busy-tick, catapulting the task(s) to
highly interactive status - while in reality they are constantly burning
CPU time.

The fix is to round down sleep_time, not to round it up.  This slightly
under-estimates the sleep time, but this is not a real problem, any task
with a sleep time in the 1 jiffy range will see timekeeping granularity
artifacts from various parts of the kernel anyway.  We could use rdtsc
to estimate the sleep time, but i think that's unnecessary overhead.

The fixups in Mike's scheduler patch (which is in -mm8) basically work
around this bug.  The patch below definitely fixes the contest-load
starvation bug, but it remains to be seen what other effects it has on
interactivity.  In any case, this bug in the estimator is real and if
there's any other interactivity problem around then we need to deal with
it ontop of this patch.

This bug has been in the O(1) scheduler from day 1 on basically, so i'm
quite hopeful that a number of interactivity complaints are fixed by
this patch.