One Model to Critique Them All: Rewarding Agentic Tool-Use via Efficient Reasoning