Small fixes and improvements to BB-lite

• Protect against infinite gradient if the error in a bin is zero. The yields are computed as total_yields(x) + BBlite_nuisances * sqrt(total_variances(x)), where x are all the variables in the model, and BBlite_nuisances is the array of BB-lite nuisance parameters. But if total_variances(x) is zero, then the derivative of sqrt(total_variances(x)) is infinite. This adds a protection in case some bins have zero variance (but this should generally be avoided anyway).
• Keep a reference to the BB-lite variable in the model as model.bblite_vars
• Give meaningful names (using channel names) to the BB-lite variable sub-components