All in Tokens: Unifying Output Space of Visual Tasks via Soft Token