Untangling Lariats: Subgradient Following of Variationally Penalized Objectives