Multi-Agent Learning with Heterogeneous Linear Contextual Bandits