Think-Then-React: Towards Unconstrained Human Action-to-Reaction Generation