Function group_norm_forward 

pub fn group_norm_forward<T: Triton, D: Float, const BLOCK_NL: i32>(
    x_ptr: T::Pointer<D>,
    y_ptr: T::Pointer<D>,
    weight_ptr: T::Pointer<D>,
    bias_ptr: T::Pointer<D>,
    mean_ptr: T::Pointer<D>,
    rstd_ptr: T::Pointer<D>,
    _N: i32,
    C: i32,
    L: i32,
    G: i32,
    eps: f32,
)
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>>>,

GroupNorm training forward — saves per-(n,g) mean and rstd.

Grid: [N * G] — one CTA per (sample, group).