To delay for less than 1ms (was: 10ms) we must busy-loop, because the
kernel is jiffy-based. No way out of it.

The implementation, however, was based on gettimeofday(), which means
we couldn't really strace the hal process, as it appeared to
only call gettimeofday all the time.

This code uses a local software loop, after a calibration step of
around 20ms (depends on CPU speed). Moreover, by calling usleep() for
longer-than-1ms delays it doesn't use 100% cpu power during the
initial led blinking.

Now strace on the process shows what it really is doing (besides
user-spac-only i2c accesses).
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>