* Optimize the cuda implementation of sum_op, which add two lod_tensors inplace.
test=develop

* Use eigen to add to tensors.
test=develop