Pass LoD from input to output in the operators.

[qingqing01]

Fix #4047
Fix #4151

1. Pass LoD from input to output in the operators.

   • Replace Tensor by LoDTensor in each operator. #4047 (comment)

2. 去掉InferShape中LoDTensor

   • 讨论：LoDTensor in operators. #4151

3. 没有传递LoD的Op

   • concat_op:
     • @Yancey1989 正在写的sequence_concat会处理LoD
   • gather_op, scatter_op, cond_op:
     • 可能会改变LoD，目前该Op没有处理变长序列，如果将来处理变长序列，需要处理LoD
   • mean_op:
     • 输出是一个标量，不用传递
   • split_op：
     • 有个SubSequenceOp，序列任务或许会用SubSequenceOp。目前没有传递，以后如果有需要再加。
   • sgd_op等optimizer
   • uniform_random_op, gaussian_random_op给Tensor填充值的Op
   • softmax_op
     • 会有sequence_softmax_op
   • rank loss, huber loss, trans_op, conv_2d_op