Function batch_norm_backward 

pub fn batch_norm_backward<T: Triton, D: Float, const BLOCK_N: i32>(
    dy_ptr: T::Pointer<D>,
    x_ptr: T::Pointer<D>,
    dx_ptr: T::Pointer<D>,
    weight_ptr: T::Pointer<D>,
    mean_ptr: T::Pointer<D>,
    rstd_ptr: T::Pointer<D>,
    dweight_ptr: T::Pointer<D>,
    dbias_ptr: T::Pointer<D>,
    N: i32,
    C: i32,
)
where
    T::I32Tensor: Tensor<i32, 1> + Comparison<i32, BoolTensor = T::BoolTensor>,
    T::Pointer<D>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<D>>>,

Computes gradients for BatchNorm.

Given saved mean and rstd from the forward pass:

xhat      = (x - mean) * rstd
dbias[c]  = Σ_n dy[n,c]
dweight[c]= Σ_n dy[n,c] * xhat[n,c]
dx[n,c]   = weight[c] * rstd[c] * (dy[n,c]
              - dbias[c] / N
              - xhat[n,c] * dweight[c] / N)

Uses two sequential passes over N within the same CTA to avoid storing the full xhat tensor.

Grid: [C] — one CTA per channel.