An Empirical Study of Reward Specification and Benchmark Reliability in GRPO-based LLM Unlearning