Prompt caching

Also: prefix caching

Reusing the computed KV cache for a prompt prefix that has not changed, so a long fixed preamble is processed once and skipped thereafter. Large savings in latency and cost for anything with a stable system prompt.