Text this: In Shift and In Variance: Assessing the Robustness of HAR Deep Learning Models Against Variability