Post-training for Ricursive Intelligence's Ripple-65B Preview incorporated synthetic data pipelines, long-context stages at 2 million sequence length, and reinforcement learning using truncated importance sampling