What strategies would you use to minimize the memory footprint of the Key-Value (KV) cache during long-context LLM inference on Tenstorrent?
