Deep Learning of Semi-Competing Risk Data via a New Neural Expectation-Maximization Algorithm