Unleashing Reasoning Capability of LLMs via Scalable Question Synthesis from Scratch