Unlocking Memorization in Large Language Models with Dynamic Soft Prompting