KV Cache and PagedAttention

Use this note for KV-cache sizing, prefill versus decode, paged allocation, batching, fragmentation, scheduling, and attention-kernel implications.