Use fused softmax kernel in llama attention layer

Use fused softmax kernel in llama attention layer #3584