Towards Generalizable Deepfake Detection via Clustered and Adversarial Forgery Learning