The Emergence of Essential Sparsity in Large Pre-trained Models: The Weights that Matter