Allowing some tolerance on the fairness constraint in the ThresholdOptimize is now featured in fairlearn v0.13.0: see for instance the user guide and the associated API reference).
ThresholdOptimize