Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment