Centurio: On Drivers of Multilingual Ability of Large Vision-Language Model