Learning Multi-Agent Loco-Manipulation for Long-Horizon Quadrupedal Pushing