be recomputed on upcast chunks.
TODO: use a float64 accumulator in row_norms to avoid the latter.
Nr[