Function muon_frob_norm_sq 

pub fn muon_frob_norm_sq<T: Triton, const BLOCK_SIZE: i32>(
    x_ptr: T::Pointer<f32>,
    out_ptr: T::Pointer<f32>,
    n_elements: i32,
)
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>>>,

Parallel Frobenius squared-norm via atomic reduction.

Each block sums x[block]² and atomically adds to out_ptr[0]. Caller must zero out_ptr[0] before launch. After launch: out_ptr[0] = ||X||_F².

Grid: [ceil(n_elements / BLOCK_SIZE), 1, 1].