Prompting Large Language Models with Rationale Heuristics for Knowledge-based Visual Question Answering