SelectIT: Selective Instruction Tuning for Large Language Models via Uncertainty-Aware Self-Reflection