sudoLLM: On Multi-role Alignment of Language Models