Tangential Randomization in Linear Bandits (TRAiL): Guaranteed Inference and Regret Bounds