Evaluating Multilingual, Context-Aware Guardrails: A Humanitarian LLM Use Case