What is the state of art way to do batchnorm between embedding layer and LSTM layer for padded sequences?