On the Optimal Memorization Power of ReLU Neural Networks