correct finding normalization value for relerr compuation

Before it would have taken the larger of the abs value of the new and old data of the last data set rather than the largest abs value of the new and old data of all the datasets.

This will make relative errros smaller and will be particularly noticeable in cases where the last dataset was small or zero.

This should get rid of the "Error, how did I get here" warnings (that really indicated an logic error).