Text-driven Affordance Learning from Egocentric Vision