couple optimiztion I did in this pr for the Graph_Input_matched

1. cache the seed_info_id instead of getting it every time
2. cache the arg_idxs, it should be the same for all calls so no need to calculte for every execution
3. pre-allocate the real_input instead of calling append over and over
4. move the if tensor_id == self.seed_info_id check into the if arg_idx is None since it is a uncommon case

with all of these I was able to reduce the runtime of this function from 0.3ms to 0.2ms for the llama3 8B case for vllm