Doubly-Optimistic Play for Safe Linear Bandits