An Exponential Speedup in Parallel Running Time for Submodular Maximization without Loss in Approximation