How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM?