The bug is that we do not clip the sequences to n_steps, meaning that when
we revert them for computing the gradient we mess up the gradient.