Ranking LLM-Generated Loop Invariants for Program Verification