Provably Reliable Conformal Prediction Sets in the Presence of Data Poisoning