Scalable Ensembling For Mitigating Reward Overoptimisation