Toward Generalist Anomaly Detection via In-context Residual Learning with Few-shot Sample Prompts