Excess contention in ManagedChannelImpl.exitIdleMode

[carl-mastrangelo]

Each RPC needs a transport, and thus calls ManagedChannelImpl's transport provider. This tries to make the transport wake up, but does so under a lock. Profiling reveals a high amount of contention on this lock. If possible we should look at eliding the lock and using a volatile read to get the loadbalancer out of it.

cc @zhangkun83

[carl-mastrangelo]

FYI, adding a local fix shows a significant improvement, so this should be treated more importantly than the other contention issues, but less than #2118