Catching Prompt Regressions Before They Ship: Semantic Diffing for LLM Workflows