Function gradient_centralization 

pub fn gradient_centralization<A, D>(
    gradients: &mut Array<A, D>,
) -> &mut Array<A, D>
where
    A: Float + ScalarOperand,
    D: Dimension,

Compute gradient centralization