[Bug Fix]clear preTopology's output while cloneCells

[qiuxin2012]

What changes were proposed in this pull request?

Recurrent cost too many memories, find it's caused by cloneCells.
clear preTopology's output while cloneCells.

With this fix, 9-depth deepspeech2 model forward a input tensor of size (1, 1, 13, 3000) cost only 1GB memory.
Without this fix, the model will cost about 80GB memory.

How was this patch tested?

unit tests, manual tests

[yiheng]

LGTM. Jenkins result?

[qiuxin2012]

@yiheng failed, some times preTopology is null.

[qiuxin2012]

jenkins passed

[zhichao-li]

@qiuxin2012 @yiheng
Would it be better if we set preTopology of a cell to be null in add method as there's weights, grad or other stuff within a preTopology other than output

  override def add(module: AbstractModule[_ <: Activity, _ <: Activity, T]): this.type = {
    if (topology.preTopology != null) {
      val tmp = topology.preTopology.cloneModule()
      topology.preTopology = null 
      preTopology = TimeDistributed(tmp, maskZero = maskZero)
    }