Risk-Sensitive Q-Learning in Continuous Time with Application to Dynamic Portfolio Selection