Function w8a16_gemm_t 

pub fn w8a16_gemm_t(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    input: DevicePtr,
    weight_t: DevicePtr,
    block_scale_t: DevicePtr,
    output: DevicePtr,
    m: u32,
    n: u32,
    k: u32,
    stream: u64,
) -> Result<()>

W8A16 Transposed GEMM: C[M,N] = A[M,K] @ dequant(B_t[K,N]) with coalesced reads.

Uses transposed FP8 weights B_t[K,N] and block_scale_t[K/128, N/128] for coalesced N-dimension reads. ~14x faster than non-transposed w8a16_gemm at long M. Grid: (ceil(N/64), ceil(M/64), 1) Block: (128, 1, 1)