ZML’s new LLMD software cuts inference costs across AI chips. Here’s why it matters and how it works—backed by Yann LeCun.