In-Context Alignment: Chat with Vanilla Language Models Before Fine-Tuning