Privacy Amplification by Structured Subsampling for Deep Differentially Private Time Series Forecasting