Zero-shot Meta-learning for Tabular Prediction Tasks with Adversarially Pre-trained Transformer