Dilated Convolution with Learnable Spacings: beyond bilinear interpolation