Invariance Makes LLM Unlearning Resilient Even to Unanticipated Downstream Fine-Tuning | Read Paper on Bytez