Making the multiplication in normalise_residual() produce a 32-bit result just

to be on the safe side (not that I observed any overflow, but I can't prove
it can't happen).