Text this: Transformer-based multi-task learning for table tennis motion feature recognition