Function for_each_row 

pub fn for_each_row<F>(
    output: &mut [f32],
    n_rows: usize,
    n_cols: usize,
    row_fn: F,
)
where
    F: Fn(usize, &mut f32) + Send + Sync,

One output row per slot; parallel when should_parallelize says the matvec is big enough to pay for the fork-join. Rows are never split.