Function moe_gate_topk_fused 

pub fn moe_gate_topk_fused(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    input: DevicePtr,
    gate_weight: &QuantizedWeight,
    expert_indices: DevicePtr,
    expert_weights: DevicePtr,
    num_experts: u32,
    k: u32,
    top_k: u32,
    normalize: u32,
    stream: u64,
) -> Result<()>

Fused gate GEMV + topK softmax for M=1 decode.

Single kernel that computes gate[num_experts] = A[K] @ B_gate[num_experts, K] then extracts top-K indices + softmax weights. Saves 1 launch vs separate gate GEMV + topK kernels.

Grid: (1, 1, 1) Block: (256, 1, 1) — single CTA, uses shared memory reduction