Use pairwise summation inside `sum_of_errors_evaluator`

[morinim]

Pairwise summation is more accurate and can be executed in parallel.

It requires a std::ranges::sized_range and std::ranges::random_access_range.

[morinim]

The simple online mean is generally better than naive $sum \over n$ with a left-to-right sum, but it is still sequential and still accumulates rounding error over time.

A rough intuition is:

• naive running sum → error can grow roughly like $O(n \epsilon)$;
• online mean update → also sequential, typically still in that broad class;
• pairwise reduction → error tends to grow more like $O(log(n \epsilon))$

So pairwise is usually more stable asymptotically.

That said, asymptotics can exaggerate the practical difference. In many real workloads the simple online mean is already very respectable and is usually faster.

Also, we already have parallelism at a higher level, introducing extra parallelism inside pairwise summation is often not worth the added complexity.