[Pitch] Balanced binary reduction (stdlib addition)

[Nevin]

@scanon

For the usual Kahan summation, I would expect that you can write it directly on AdditiveArithmetic and the compiler will optimize away the computation of excess when the element type is an integer (since it is trivially always zero), so it probably doesn't need a customization point at all. For Neumaier's improved Kahan summation, which you appear to be using here, it may be slightly harder to achieve this, but it's probably still possible with a little bit of care.

I'm not opposed to adding new customization points, but the bar to adding them on a protocol like AdditiveArithmetic should be high.

I would want a different algorithm for signed integers, as described here in that same thread, to avoid spurious overflow.