Learning Time-Continuous Dynamics Models With Gaussian-Process-Based Gradient Matching