Inference under covariate-adaptive randomization with multiple treatments