LLM Unlearning on Noisy Forget Sets: A Study of Incomplete, Rewritten, and Watermarked Data