UniFormer: Unifying Convolution and Self-attention for Visual Recognition