Cartridges and STILL simplify KV-cache benchmarking

A public, single-GPU code release reproduces two recent long-context inference ideas: Cartridges for corpus-specific compressed KV caches and STILL for reusable neural KV-cache compaction. The repos emphasize runnable benchmarks, readable implementations, and direct comparisons against full-context inference, truncation, and Cartridges.