Function moe_expert_gemv_silu_down_2x 

pub fn moe_expert_gemv_silu_down_2x(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    gate_out: DevicePtr,
    up_out: DevicePtr,
    packed_ptrs: DevicePtr,
    scale_ptrs: DevicePtr,
    scale2_vals: DevicePtr,
    output: DevicePtr,
    expert_indices: DevicePtr,
    n: u32,
    k: u32,
    top_k: u32,
    stream: u64,
) -> Result<()>

Register-tiled fused SiLU+down expert GEMV: 2 output rows per thread.

Same as silu_down but each thread computes 2 adjacent output rows, reusing the SiLU(gate)*up activation from registers. Doubles weight reads per iteration for better LPDDR5X bandwidth utilization.

Grid: (ceil(N/8), top_k, 1) Block: (128, 1, 1)