[RFC] Tail-Optimized LRU (T-LRU): Reducing Tail Latency via Conversation-Aware KV Cache Eviction