Learning Clinical Representations through Ontology-Aware Contrastive Pretraining and Cross-Modal Distillation