SERVER-139309ClosedFixed
make the CMS channel manager concurrency-safe
We use an internal synchronized state which is copied when callers make a CMS RPC. Channel rotations are guarded so that if two concurrent RPCs fail and want to rotate the channels, only one of them will do so and the other will detect the stale rotate and bail out. Channels are initialized lazily. Aside from being a useful way to spread out connections, this is the same behaviour as before and several tests depend on it.