Overhaul reproducibility error propagation by using d18O residuals

[nielsjdewinter]

A potentially more realistic and stable uncertainty of reproducibility can be obtained from the residuals of d18O in individual models, which are already calculated in the run_model function.
Currently, reproducibility is based on:

1. propagated errors on the input (data uncertainty)
2. Reproducibility errors on the temperature and age results between windows (model uncertainty)
   The model uncertainty estimate could be improved by considering d18O residuals, since d18O is the actual fitted parameter and contains more direct information about the model success.