Function fp8_fp8_gemm_n128 

pub fn fp8_fp8_gemm_n128(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    a_fp8: DevicePtr,
    b_fp8: DevicePtr,
    output: DevicePtr,
    m: u32,
    n: u32,
    k: u32,
    stream: u64,
) -> Result<()>

FP8×FP8 GEMM: A [M, K] FP8 × B [N, K] FP8 → C [M, N] BF16.

Both A (activations) and B (weights) are pre-converted FP8 E4M3. No BF16→FP8 conversion in inner loop — pure MMA throughput. Grid: (ceil(N/128), ceil(M/64)) Block: (128, 1, 1)