Function mla_q_rope_extract_batched 

pub fn mla_q_rope_extract_batched(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    q_full: DevicePtr,
    q_rope_out: DevicePtr,
    num_tokens: u32,
    nq: u32,
    hd: u32,
    nope: u32,
    rope: u32,
    q_dim: u32,
    stream: u64,
) -> Result<()>

Batched Q rope extract: [N, nq, hd] → [N, nq, rope] at offset nope per head. 1 kernel replaces N*nq D2D copies per layer.