Fix ThreadLocal races

Fix ThreadLocal races

Summary:
I misread the ThreadLocal docs, thread destruction functions do *not* grab the accessAllTHreads_ lock
unless you use *strict* mode, and even then, it is only a read lock.

Easy enough to make the thread-destruction global bits to be atomic / use folly::Synchronized.

Reviewed By: yfeldblum

Differential Revision: D6592905

fbshipit-source-id: 4ae600dff4c8c04751483a452ca7c07ef3f26380