An Efficient Closed-Form Solution to Full Visual-Inertial State Initialization