Scatter data to evaluators in a more fine-grained way

[jsjason]

When we start to consider DNN model partitioning, we need to come up with a data distribution scheme that can correctly divide and scatter data cross evaluators. For example, evaluators that have partitions of the same model (not replicas) need to receive a scatter of the same training example, whereas evaluators of different model replicas need to receive totally different training examples.