Finite-Time Analysis for Conflict-Avoidant Multi-Task Reinforcement Learning