Can Medical Vision-Language Pre-training Succeed with Purely Synthetic Data?