Learning Paths for Dynamic Measure Transport: A Control Perspective