fix handling of non-physical data when calculating staterrors

[lukasheinrich]

Description

Based on the way the data is collected using MC, one can get into situations where the a ends up having zero counts but also adds a statsys error (or similar) of zero as well. Currently we do assign a nuisance parameter for this, but then crash because this check

https://github.com/diana-hep/pyhf/blob/master/pyhf/modifiers/staterror.py#L122

fails for the bins in the channel in which the count is zero.

We can

• preprocess the data and not allocate a nuisance parameter for it (this could be a generic pruning step of removing parts from the model that do not make sense)
• add negligible values to it

probably the first is more correct, but we will need to see how

Expected Behavior

zero counts happening just because of limited MC stat should be handled gracefully

Actual Behavior

pyhf carshes

Steps to Reproduce

create a model in which a staterror is added but one bin has zero counts and zero uncertainty (summed)

Checklist

• Run git fetch to get the most up to date version of master
• Searched through existing Issues to confirm this is not a duplicate issue
• Filled out the Description, Expected Behavior, Actual Behavior, and Steps to Reproduce sections above or have edited/removed them in a way that fully describes the issue