See https://github.com/orgs/scipp/discussions/3595#discussioncomment-11222220

Investigate possible performance benefits of using multiprocessing to speed up the scipp.curve_fit routine.

Questions:

1. Do we need any size thresholds for when to use the parallel version?
2. In case of serialization problems, do we fail gracefully or just fail? Drawback of failing gracefully is increased code complexity to handle an unusual edge case.