Echo: Decoupling Inference and Training for Large-Scale RL Alignment on Heterogeneous Swarms