Finite-Time Error Analysis of Soft Q-Learning: Switching System Approach