Smoothing the Landscape Boosts the Signal for SGD: Optimal Sample Complexity for Learning Single Index Models