Optimising longitudinal and lateral calorimeter granularity for software compensation in hadronic showers using deep neural networks