A Self-Pruning Transformer: Extreme KV-Cache Compression w/Universal Attention