Leveraging Complementary Embeddings for Replay Selection in Continual Learning with Small Buffers