How we cut LiteLLM's Redis round trips per request by 64%

A governed request to the LiteLLM AI Gateway waited on Redis 22 times. It now waits 8 times: one pipeline per Redis backend before the model call, one after.