Class-Dependent Perturbation Effects in Evaluating Time Series Attributions