Reachability-Aware Reinforcement Learning for Collision Avoidance in Human-Machine Shared Control