Collaborative Linear Bandits with Adversarial Agents: Near-Optimal Regret Bounds