Times package:compensated

times a b k computes k x y such that

x + y = a * b
x = fl(a * b)

Which is to say that x is the floating point image of (a * b) and y stores the residual error term. This could be nicer if we had access to a hardware fused multiply-add.