Function moe_unpermute_reduce_indexed 

pub fn moe_unpermute_reduce_indexed(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    expert_output: DevicePtr,
    output: DevicePtr,
    token_to_perm: DevicePtr,
    topk_weights: DevicePtr,
    hidden_size: u32,
    num_tokens: u32,
    topk: u32,
    stream: u64,
) -> Result<()>

Unpermute + weighted reduce with pre-built reverse map.

Grid: (num_tokens, 1, 1) Block: (256, 1, 1)