Understanding Deep Neural Function Approximation in Reinforcement Learning via $\epsilon$-Greedy Exploration