The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning