Beyond Precision: Importance-Aware Recall for Factuality Evaluation in Long-Form LLM Generation