Knockoff-Guided Feature Selection via A Single Pre-trained Reinforced Agent