Self-generated RL curricula

Training loops where the model proposes its own tasks and difficulty schedule instead of consuming a fixed human-curated task set.