PREMERE: Meta-Reweighting via Self-Ensembling for Point-of-Interest Recommendation