AREX: Towards a Recursively Self-Improving Agent for Deep Research