The Good, the Bad, and the Sampled: a No-Regret Approach to Safe Online Classification