$\text{Alpha}^2$: Discovering Logical Formulaic Alphas using Deep Reinforcement Learning