Sparse attention at million-token context

A linear-cost attention variant that holds quality past a million tokens.