GoT-R1: Unleashing Reasoning Capability of MLLM for Visual Generation with Reinforcement Learning