SpeedLLM: An FPGA Co-Design of Large Language Model Inference Accelerator