Return of the Encoder: Maximizing Parameter Efficiency for SLMs