LoRA Subtraction for Drift-Resistant Space in Exemplar-Free Continual Learning