we buzzloop calling gettimeofday() until we see a big jump.
old code would take the difference between the immediately previous return value.
Now, compute the difference based on the clock value as of right
after the previous big jump.
An alpha pc164 at 500MHz can do about 700 calls to gettimeofday() in
one clock tick, each of which increments tv_usec by 1 -- so the "big
jump" at the end is only ~270us rather than the real clock tick of
976us, which yields a precision value which is overoptimistic by a
factor of ~4. The corrected code now yields the exact tick value
(which is correct since NetBSD/alpha doesn't have a precise
microtime).
