Fixed

• Make sure the default bucket is large enough with word based batching when the source is longer than the target (Previously
  there was an edge case where the memory usage was sub-optimal with word based batching and longer source than target sentences).