Towards AI-Safety-by-Design: A Taxonomy of Runtime Guardrails in Foundation Model based Systems