Optimizing Facial Landmark Estimation for Embedded Systems Through Iterative Autolabeling and Model Pruning