Strefer: Empowering Video LLMs with Space-Time Referring and Reasoning via Synthetic Instruction Data