Detect, Unlearn, Restore: Defending Text Summarization Models Against Data Poisoning