Multi-agent Reinforcement Learning-based Capacity Planning for On-demand Vehicular Fog Computing