EG-VQA: Benchmarking Verifiable Video Question Answering with Grounded Temporal Evidence