Evaluating Zero-Shot Long-Context LLM Compression