Rates of Convergence of Performance Gradient Estimates Using Function Approximation and Bias in Reinforcement Learning