Function moe_w4a16_grouped_gemm_ptrtable_k64_n128 

pub fn moe_w4a16_grouped_gemm_ptrtable_k64_n128(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    a: DevicePtr,
    b_packed_ptrs: DevicePtr,
    b_scale_ptrs: DevicePtr,
    scale2_vals: DevicePtr,
    c: DevicePtr,
    expert_offsets: DevicePtr,
    sorted_token_ids: DevicePtr,
    num_experts: u32,
    n_out: u32,
    k: u32,
    max_m_tiles: u32,
    stream: u64,
) -> Result<()>

K64 down GEMM: K_STEP_T=64 eliminates pipeline stall (compute=128 cycles > load ~100 cycles). Use when K=inter (512 for 35B) — 8 K-steps vs 16 with K32.

Grid: (ceil(n_out/128), max_m_tiles, num_experts) Block: (128, 1, 1)