Temporal Preferences in Language Models for Long-Horizon Assistance