DeepSeek drops input cache price to 1/10th

Long chats are mostly cache. For coding agents this is like >90%.