Tau-FPL: Tolerance-Constrained Learning in Linear Time