Atomic Consistency Preference Optimization for Long-Form Question Answering