Mitigating silent failures in toxicity prediction: a conformalized heteroscedastic Bayesian framework | Synapse