Beyond Post-Hoc Temperature Scaling: Bilevel Optimization for LLM Calibration