Function quant_gemv 

pub fn quant_gemv(
    gpu: &dyn GpuBackend,
    gemv_nvfp4: KernelHandle,
    gemv_fp8: KernelHandle,
    gemv_dense: KernelHandle,
    input: DevicePtr,
    weight: &QuantWeight,
    output: DevicePtr,
    n: u32,
    k: u32,
    stream: u64,
) -> Result<()>

Unified GEMV dispatch: select kernel based on weight quantization format.

Eliminates cascading if/else chains in layer forward methods. The enum branch (~1 cycle) is negligible vs GPU kernel launch overhead (~5μs).