Singular Value Penalization and Semantic Data Augmentation for Fully Test-Time Adaptation