Function moe_expert_silu_down_shared_fp8_batch3_t 

pub fn moe_expert_silu_down_shared_fp8_batch3_t(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    gate_out: DevicePtr,
    up_out: DevicePtr,
    weight_t_ptrs: DevicePtr,
    block_scale_t_ptrs: DevicePtr,
    output: DevicePtr,
    expert_indices: DevicePtr,
    sh_gate_in: DevicePtr,
    sh_up_in: DevicePtr,
    sh_down_t: &Fp8Weight,
    sh_down_out: DevicePtr,
    n: u32,
    k: u32,
    top_k: u32,
    stream: u64,
) -> Result<()>

FP8 fused SiLU+down GEMV (transposed). K=3 batch.