Benchmarking Counterfactual Interpretability in Deep Learning Models for Time Series Classification