kv-cache-size 0.1.1

Exact KV-cache arithmetic for transformer inference: bytes per token, total cache bytes, and the context length that fits a memory budget.