Function w4a16_gemv_qkvz 

pub fn w4a16_gemv_qkvz(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    input: DevicePtr,
    weight: &QuantizedWeight,
    output: DevicePtr,
    n: u32,
    k: u32,
    num_groups: u32,
    head_k_dim: u32,
    vheads_per_group: u32,
    head_v_dim: u32,
    stream: u64,
) -> Result<()>

W4A16 GEMV with inline QKVZ deinterleave on output write.

Same as w4a16_gemv but writes to deinterleaved output locations, eliminating the separate deinterleave_qkvz kernel.

Kernel: w4a16_gemv_qkvz(A, B, S, s2, C, N, K, ng, kd, vpg, vd) Grid: (ceil(N/4), 1, 1) Block: (256, 1, 1)