What is the KV-cache, why is it critical for optimizing LLM generation speed, and how does PagedAttention solve KV-cache fragmentation on AMD GPUs?
