FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification