max_pool2d_with_indices_cuda and max_pool2d_with_indices_backward_cuda should have channel last optimized kernels ( for https://github.com/pytorch/pytorch/pull/24872 )