trigger-rebuild leaves outdated kb in memory of sibling processes

[jordanstephens]

Receiving a request to the /trigger-rebuild endpoint and updating the in-process memory directly still leaves the sibling processes unaware of the new knowledge base. One solution that occurs to me is having the /trigger-rebuild endpoint immediately write the new knowledge base data to the on-disk cache. When we're handing an inference request, we can check the mtime of the cache files and if they are newer that the last time we read them, we can read them again before proceeding. We would still want to keep the lock on the read and write code paths.