Adaptive Sparsity Level during Training for Efficient Time Series Forecasting with Transformers