Resolve pernicious race between destructor and thread functions by insisting that destructor is invoked after the class has gone back to being a harmless single-threaded thing.

This reveals an issue that needs to be solved:
D0713 14:48:28.049861213   14503 server.c:704]               Waiting for 44 channels and 0/1 listeners to be destroyed before shutting down server

This is for the same reasons as ParseJson, that is so that we can limit
the scope of the proto namespace differences between internal and
external.

stability. I believe that this change was unneeded and actually
may hurt matters by holding a lock for too long.
Partially undoes #5444 , particular commit e7042b5d

Server was continuing to make requests for new calls forever, which were
starving out the shutdown sequence. Change order and win.

C core automatically shares subchannels between channels. Therefore our
multiple channel performance tests were really testing single channel
performance.

input

distributions since we don't use them and it's not sensible
to add them