Efficient Accelerator/Network Co-Search With Circular Greedy Reinforcement Learning