Improving In-Context Few-Shot Learning via Self-Supervised Training