[Discussion] Why is it assumed that dynamic activation reordering using the Hessian diagonal must improve performance (accuracy, perplexity, etc.)?