prime-rl: scale inference pods mid-training → auto-discovered → weight auto-sync → rollouts auto-route