Game-RL: Synthesizing Multimodal Verifiable Game Data to Boost VLMs' General Reasoning