Chunks as Arms: Multi-Armed Bandit-Guided Sampling for Long-Context LLM Preference Optimization