Reverted commit D3755446

Reverted commit D3755446

Summary:
Move ThreadLocal object destruction to occur under the lock to avoid races.
This causes a few cascading changes - the Tag lock needs to be a recursive_mutex so
constructing a new object while destroying another st. Also, forking requires
a new mutex to avoid deadlocking on accessing a recursive_mutex across a fork()

Reviewed By: andriigrynenko

Differential Revision: D3755446

fbshipit-source-id: f1f1f92175eb39e77aaa2add6915e5c9bb68d0fb