Almost Optimal Batch-Regret Tradeoff for Batch Linear Contextual Bandits