Function w4a16_gemv_qg 

pub fn w4a16_gemv_qg(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    input: DevicePtr,
    weight: &QuantizedWeight,
    output: DevicePtr,
    n: u32,
    k: u32,
    num_heads: u32,
    head_dim: u32,
    stream: u64,
) -> Result<()>

W4A16 GEMV with inline Q/Gate deinterleave on output write.

Same as w4a16_gemv but writes Q and Gate to deinterleaved positions, eliminating the separate deinterleave_qg kernel (12 graph nodes saved).

Kernel: w4a16_gemv_qg(A, B, S, s2, C, N, K, num_heads, head_dim) Grid: (ceil(N/4), 1, 1) Block: (256, 1, 1)