Sequoia: Scalable, Robust, and Hardware-aware Speculative Decoding