Optimal Block-Level Draft Verification for Accelerating Speculative Decoding