BabyHGRN: Exploring RNNs for Sample-Efficient Training of Language Models