Risk-Averse Policy Gradient for Tail Risk Optimization Using Extreme Value Theory