Combining channel-wise joint attention and temporal attention in graph convolutional networks for skeleton-based action recognition