Selective layer-wise cleansing: A knowledge-preserving defense against backdoor attacks in LoRA-tuned models for natural language understanding