OST: Improving Generalization of DeepFake Detection via One-Shot Test-Time Training