Tracking the accuracy/uncertainty of operations

[kreczko]

Recent adjustments to the codebase unearthed some uncertainties that we need to find a way to keep track off.

The current example is ak.sum vs np.sum where the result is dependent on the parameters used (see scikit-hep/awkward#1241).

I've found some old numpy discussions but also a related project - accupy.

In the same spirit as we teach students how to do error-propagation for experiments they do, we should find a way to at least estimate the uncertainties of operations performed.

We might want to have a look at unumpy

A --with-error-propagation option would be the best possible outcome.

In the meantime we should do our best to make people aware of the numerical limitations.