Estimator-Coupled Reinforcement Learning for Robust Purely Tactile In-Hand Manipulation