CRANE: Reasoning with Constrained LLM Generation