Provide a configurable option to stream all completed_requests to disk.

Currently completed_requests is a per-node in-memory cache only.  When a node restarts, the cache content is lost.   The idea behind this improvement is to provide a mechanism that saves all completed requests to files on the local node in an effort to have a more complete picture of activity over a longer period than can be accomplished with an in-memory only solution.

Mindful of the potential performance impact, this must be an opt-in feature that can be simply enabled and disabled.   It is also expected that the writing to disk will lag behind requests so it won't entirely eliminate the potential to "lose" requests in the event of an outage.  Obviously it won't track requests which have yet to finish - it isn't a log of received requests.

It is intended to only capture requests that would also be recorded in the in-memory cache; that is, the existing qualifier mechanism determines if a request will be recorded and therefore configures the request capture aspect of this improvement.

Configuration for the improvement will amount to a size (completed-stream-size, in MiB) for the total amount of history to retain - this will be the approximate disk space used by the feature, which will be managed by automatic removal of the oldest data. (Given there will be parallel streams writing to independent files (performance), "oldest" will not be a strict constraint - "generally oldest" may be more accurate.)