Toward Low-Bit Neural Network Training Accelerator by Dynamic Group Accumulation