When we are trying to covert the existing float32 tensor to a float16 one, we also need to copy the LoD information.