Reduce Kafka repartitioning delays when client sidecar joins an application mesh

[dgrove-oss]

Optimize our Kafka partitioning/rebalancing algorithm to reduce (or ideally eliminate) the stop-the-world pause that occurs when a a client-only sidecar joins the application mesh. One common scenario for this is when a kar invoke, kar rest, or kar get -mr command is run by a developer to interact with their application.

We know at start up whether a particular sidecar invocation is going to be a short-lived client (executing a single command), a longer-lived client (kar run without a -actors or -service argument) or a server (kar run with -actors and/or -service).