Speeding up Speculative Decoding via Approximate Verification