On the inapproximability of M/G/K: why two moments of job size distribution are not enough