Rethinking FID: Towards a Better Evaluation Metric for Image Generation