CARD: Cache-Assisted Parallel Speculative Decoding for Efficient Large Language Model Inference.