Veri-R1: Toward Precise and Faithful Claim Verification via Online Reinforcement Learning