AVA: Towards Agentic Video Analytics with Vision Language Models