Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models