Right now, we have different formulas to compute relative/absolute errors, and reporting the worst offenders. Some are in verify_grad, others in values_eq_approx, probably some are in diverse tests.

We should centralize and harmonize that.