2. Remove num_seq arguments.
3. Refine CUDA kernel of ScaleLoDTensorFunctor.
4. Change max_relative_error of gradient unitest to 0.007