Nearly Tight Bounds for Cross-Learning Contextual Bandits with Graphical Feedback