Toward Training Superintelligent Software Agents Through Self-Play SWE-RL (Meta)