Faster Text Generation with Self-Speculative Decoding