Function multi_margin_loss_backward 

pub fn multi_margin_loss_backward<T: Triton, const BLOCK_SIZE: i32>(
    dy_ptr: T::Pointer<f32>,
    input_ptr: T::Pointer<f32>,
    targets_ptr: T::Pointer<i32>,
    dx_ptr: T::Pointer<f32>,
    _n_rows: i32,
    n_cols: i32,
    margin: f32,
)
where
    T::I32Tensor: Tensor<i32, 1> + Comparison<i32, BoolTensor = T::BoolTensor>,
    T::Pointer<f32>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<f32>>> + AddOffsets<i32, 1, T::Tensor<i32>, Output = T::Tensor<T::Pointer<f32>>>,
    T::Pointer<i32>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<i32>>>,
    T::Tensor<i32>: Tensor<i32, 1>,

Multi-margin loss backward (per-row).

Grid: [n_rows, 1, 1]. BLOCK_SIZE must equal next_power_of_two(n_cols).

Two-step write:

1. Store (dy / n_cols) * active_j for all j in the row.
2. Atomic-add correction at target: -(dy / n_cols) * sum(active_all).