Fix bias in ordering SRV RR results by weight.
In lib/roken/resolve.c, we find rk_dns_srv_order() which re-orders
the results of an SRV RR lookup by the algorithm in RFC2782. We
fix a bias in the random weight sorting by changing the order of
operations when selecting rnd. rnd should be a non-zero random
number less than the sum of the weights at a particular priority,
but zero was included as a legitimate output thus biasing the
selection process. rk_random() % sum is still biased as a 32
bit int modulo a number which doesn't divide 2^32 does not have
a uniform distribution, but the bias should be small enough to
live with for our purposes here.