When Gradients Collide: Failure Modes of Multi-Objective Prompt Optimization for LLM Judges