The Optimal Token Baseline: Variance Reduction for Long-Horizon LLM-RL