Add compensated reflection coefficient routine

[RhysU]

Leftover TODOs from #3 that are feature-related. Namely:

• Add an ar::welford_inner_product-like routine computing a compensated reflection coefficient to be sure accumulation errors on long signals aren't killing us.
• Compare the before/after results for issue3.dat in terms of the complex pole locations.
• Move to the compensated variant if desirable.

Starting points are commented out in ar::negative_half_reflection_coefficient within ar.hpp and should be wholly contained there.

[RhysU]

Welford turns out to be a dumb approach. Kahan is much cleaner.

vimdiff on the different with/without Kahan summation possibilities for Nick's quarrelsome issue3.dat. Upper result is with. Lower result is without.

Conclusion is that the robust accumulation matters.