Working Memory Constraints Scaffold Learning in Transformers under Data Scarcity