Function fp8_gemm_n128_m128 

pub fn fp8_gemm_n128_m128(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    input: DevicePtr,
    b_fp8: DevicePtr,
    output: DevicePtr,
    m: u32,
    n: u32,
    k: u32,
    stream: u64,
) -> Result<()>

M128 variant of fp8_gemm_n128: halves B re-reads for large M (ISL > 128).

Each CTA covers 128 rows of A, loading B once for both 64-row halves. ~2× speedup on out_proj (K=value_dim, N=h) at ISL≥128.

Grid: (ceil(N/128), ceil(M/128), 1) Block: (128, 1, 1)