Privacy Amplification via Compression: Achieving the Optimal Privacy-Accuracy-Communication Trade-offs in Distributed Mean Estimation