Function deinterleave_qg_split 

pub fn deinterleave_qg_split(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    data: DevicePtr,
    q_out: DevicePtr,
    num_tokens: u32,
    num_heads: u32,
    head_dim: u32,
    stride: u32,
    stream: u64,
) -> Result<()>

Deinterleave Q/Gate with split output — Q to separate contiguous buffer.

Same as deinterleave_qg but writes Q to q_out (contiguous [N, q_dim]) instead of in-place. Gate is still written back to data in-place. Eliminates the per-token D2D copy loop.