Removing Geometric Bias in One-Class Anomaly Detection with Adaptive Feature Perturbation