For a 7B Llama-style model (32 layers, hidden size 4096, fp16), roughly how big is the KV cache per token?