improve the random stream itself (it doesn't), but to introduce noise in the arc4random calling pattern. Thanks to matthew@ who pointed out bias in a previous diff, ok deraadt@ matthew@