Summary:
We would like to guarantee that after `folly::ThreadLocal<>` dtor
returns no per-thread instances are longer alive.  Currently this is not a case:

* T1 is excuting `StaticMetaBase::onThreadExit()`, it acquired all per-thread
  instances and erased them from meta under `accessAllThreadsLock_`.
* T2 destroys `folly::ThreadLocal<>`, it executes `StaticMetaBase::destroy()`,
  collects all per-thread instances (thus missing the ones being destroyed by
  T1), destroys them and returns.
* T1 executes dtor of per-thread instances, after parent `folly::ThreadLocal<>`
  dtor is finished.

Reviewed By: ot

Differential Revision: D4109820

fbshipit-source-id: d547b8cc77c9871126538c38644c2e98ddccf220