Convergence of synchronous reinforcement learning with linear function approximation