Function matmul_vec 

pub fn matmul_vec(
    output: &mut [f32],
    input: &[f32],
    weight: &[f32],
    k: usize,
    n: usize,
)

Optimized matrix-vector multiply using NEON dot products.

For single-token inference (m=1), computes dot products between the input vector and each weight row.

Weight layout: [n, k] (row-major), so weight row j is at offset j*k.