Unmemorization in Large Language Models via Self-Distillation and Deliberate Imagination