Performant implementation of reduction operations that drop pixel grouping

[SimonHeybrock]

From scipp/essdiffraction#71 (comment):

Well, the larger issue is that we have to concat first, which in itself is avoiding other performance issues. Making many->many mapping is a long-standing issues and has seen a couple improvements in the past, but is apparently not fully solved (See scipp/scipp#1846).

What could probably be done in Scipp is bypass the single-threading issue, but it also required some thought (not sure there aren't any subtleties yet), but it is slightly odd to put an optimization for a solution that bypasses another problem.

Note the similar code exits in ESSsans.

[SimonHeybrock]

The following code in ESSsans shows that there is a slightly wider issue that we are likely going to encounter also in other techniques: https://github.com/scipp/esssans/blob/a186d7eb7c57d2b367b097596b23b9ce425c0f35/src/ess/sans/i_of_q.py#L202-L259

Trying to come up with a more unified solution with better UX on the Scipp side...