Not-a-Bandit: Provably No-Regret Drafter Selection in Speculative Decoding for LLMs