Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning