DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control