Generating Synergistic Formulaic Alpha Collections via Reinforcement Learning