Function mla_fused_prefill 

pub fn mla_fused_prefill(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    q_full: DevicePtr,
    q_rope: DevicePtr,
    kv_latent: DevicePtr,
    k_rope: DevicePtr,
    w_uk: DevicePtr,
    w_uv: DevicePtr,
    v_out: DevicePtr,
    k_cache_out: DevicePtr,
    v_cache_out: DevicePtr,
    seq_len: u32,
    nq: u32,
    nope: u32,
    rope: u32,
    kv_lora: u32,
    v_dim: u32,
    hd: u32,
    num_kv_heads: u32,
    inv_sqrt_d: f32,
    stream: u64,
) -> Result<()>

Fused MLA prefill: Q_absorption + attention + V_extraction in one kernel. Grid: (num_heads, seq_len, 1) Block: (256, 1, 1)