In a recent code change, the logic to load the migration recovery info at start() time has been moved out of the section protected by the mutex.

This might be cause the recovery info op to fail if BalancerCommandsScheduler::start() gets invoked while a prior stop() is still being executed.