RLHF from Heterogeneous Feedback via Personalization and Preference Aggregation