STVG-R1: Incentivizing Instance-Level Reasoning and Grounding in Videos via Reinforcement Learning