Cyclical Temporal Encoding and Hybrid Deep Ensembles for Multistep Energy Forecasting