Fast and Data Efficient Reinforcement Learning from Pixels via Non-Parametric Value Approximation