Towards Reliable Zero Shot Classification in Self-Supervised Models with Conformal Prediction